pygil_ensure() was taking pyexitedMu and then GIL, which can deadlock
via AB BA scenario if pygil_ensure() is also called in another thread
with GIL already held.

-> Fix it by always taking GIL with pyexitedMu released.

Without the fix the test added always hangs with thread backend, e.g.
like this:

    golang/pyx/runtime_test.py::test_pyx_pyfunc_vs_gil_deadlock
    STUCK
    terminate called after throwing an instance of 'golang::PanicError'
      what():  STUCK
    Fatal Python error: Aborted

The bug was introduced in b073f6df (time: Move/Port timers to C++/Pyx nogil).

Commit b073f6df (time: Move/Port timers to C++/Pyx nogil) added C++
level PyFunc as a way to call Python function from under nogil code.
We will soon need this functionality for sync.WorkGroup too, so move it
into shared place.

We move into regular library - not a pyx module (e.g. golang.pyx) -
because for pyx modules linking is done at import time by specially
generated code, while we need the linker support to automatically
resolve e.g. calls to PyFunc::operator() from inside std::function.

In the next patch we are going to move PyFunc from time.pyx into shared
library. Before doing that let's provide DSO visibility macros from
libgolang.h to offload other libraries from duplicating dance on how to
mark symbols visibility properly depending on compiler/OS.

Gevent runtime uses python-level calls internally which might interfere
with current python state. For example if current python exception is
set, and e.g. go or makesema runtime call is made, the following
happens:

    golang/golang_test.py::test_pyx_runtime_vs_pyexc RuntimeError: abc

    The above exception was the direct cause of the following exception:

    SystemError: <class 'gevent.__semaphore.Semaphore'> returned a result with an error set
    Exception ignored in: 'golang.runtime._runtime_gevent._sema_alloc'
    SystemError: <class 'gevent.__semaphore.Semaphore'> returned a result with an error set
    terminate called after throwing an instance of 'golang::PanicError'
      what():  makesema: alloc failed
    Fatal Python error: Aborted

-> Fix all functions in the runtimes that work at Python level to
   save/restore Python-level exception on entry/exit.

This is mostly gevent runtime, but also a couple of non-posix fallbacks in
thread runtime.

The bug was there from day 1 of runtimes - from ce8152a2 (pyx api:
Provide sleep), f971a2a8 (pyx api: Provide go) and 69db91bf (libgolang:
Add internal semaphores).

In the next patch we will fix runtime functions that work at Python
level to preserve user Python exception state. Before that, as a
preparatory step, restructure runtimes code, so that all such
Python-accessing functions are grouped together.

setuptools_dso hardcodes to use its own build_ext in its setup.
We amend that build_ext and inject our own version carefully while
setup() is run - see 7ae8c4f3 (pyx.build: Allow to combine C and C++
sources in one extension). However, currently, if user uses cmdclass =
{'build_ext': ...} in his own setup, this will be ignored and
overwritten by golang.pyx.build.build_ext . One example where this
breaks things is wendelin.core which hooks into build_ext to first
configure/build CCAN before any extension:

https://lab.nexedi.com/nexedi/wendelin.core/blob/b26ba558/setup.py#L147-153

-> Fix it by taking into account what user could put into cmdclass.

If defer was used multiple times in one function it was failing to
compile. For example added test is failing like this:

    In file included from golang/runtime/libgolang_test.cpp:22:
    golang/runtime/libgolang_test.cpp: In function ‘void __test_defer(int*)’:
    ./golang/libgolang.h:457:59: error: redeclaration of ‘golang::_deferred _defer___COUNTER__’
     #define defer(f) golang::_deferred _defer_ ## __COUNTER__ (f)
                                                               ^
    golang/runtime/libgolang_test.cpp:473:5: note: in expansion of macro ‘defer’
         defer([&]() {
         ^~~~~
    ./golang/libgolang.h:457:36: note: ‘golang::_deferred _defer___COUNTER__’ previously declared here
     #define defer(f) golang::_deferred _defer_ ## __COUNTER__ (f)
                                        ^~~~~~~
    golang/runtime/libgolang_test.cpp:470:5: note: in expansion of macro ‘defer’
         defer([&]() {
         ^~~~~

It turns out the __COUNTER__ was not expanded and they say you have to
use 3 level of macro calls. Oh well...

https://stackoverflow.com/a/27611869/9456786

std::function is frequently too long to type while func is more native to Go.

For example wendelin.core wants to accept Python-level context,
transform it into C-level context and pass to pyx/nogil code.

Provide context-related functionality that can be used directly from C++
and Pyx/nogil codes. Python-level classes and functions become small
wrappers around pyx/nogil ones.

Like with timers (b073f6df "time: Move/Port timers to C++/Pyx nogil")
and interfaces (5a99b769 "libgolang: Start providing interfaces")
memory for objects dynamically allocated on heap is managed
automatically.

Start new package that provides C++ amendments to be used by libgolang
and its users. Current functionality: dict and set that wrap
std::unordered_map and std::unordered_set into ergonomic interface.

The code originates from wendelin.core:
https://lab.nexedi.com/kirr/wendelin.core/blob/5a045ed1/wcfs/internal/wcfs_misc.h#L214-257

Pyx/nogil only.

With just one function - errors.New() - to create new error with given
text.

This will be soon used in Pyx/nogil context package to create context.canceled
and context.deadlineExceeded errors.

Pyx/nogil only for now.

- interface is analog of interface{} in Go.
- error is analog of error in Go.

For now the interfaces are implemented via classic C++ scheme via
inheritance and additional vtable field in pointed object.

In the future we might want to consider reworking that to Go-style
interfaces without predeclaring which interfaces a class implements.

Top-level documentation is TODO.

Interfaces will be soon needed to describe Context interface in C
context package.

- refptr==refptr  (e.g. to compare whether returned error object is
  something particular)
- adoptref() and newref()
- object (exposed as gobject not to be confused with builtin pyx
  "object" referring to python object)

All this will be soon used in C version of context package.

Briefly describe provided functionality instead of only referring to
https://golang.org/pkg/context and https://blog.golang.org/context.

Thinking a bit more after e82b4fab (libgolang: Objects refcounting
(initial draft)) and working on interfaces as part of context pyx -> C
move, I've came to conclusion that the better name for on-heap objects
managed by libgolang is just "object" without "ref" prefix.

-> Rename refobj -> object and amend its documentation correspondingly.

So that refptr<T> could be used as keys in e.g. dict (unordered_map) or
set (unordered_set). set<refptr> will soon be used by C version of
context package.

We already provide select for array and std::initializer_list. However
when cases are constructed at runtime dynamically both array and
initializer_list are not convenient to use.

-> Provide select() overload that accepts vector<_selcase>.

This functionality will soon be used in C version of context package.

- indicated that WorkGroup is not provided in Go version (amends
  40f3b90c "sync: Provide package documentation").
- fix typo in sync.pxd subject (amends 34b7a1f4 "golang: Expose Sema and
  Mutex as public Python and Cython/nogil API")

Every time tests are run under Python3 the following warnings are printed:

  golang/golang_test.py::test_func
    .../golang/golang_test.py:990: DeprecationWarning: inspect.getargspec() is deprecated since Python 3.0, use inspect.signature() or inspect.getfullargspec()
      assert inspect.formatargspec(*inspect.getargspec(MyClass.zzz)) == '(self, v, x=2, **kkkkwww)'

  golang/golang_test.py::test_func
    .../golang/golang_test.py:990: DeprecationWarning: `formatargspec` is deprecated since Python 3.5. Use `signature` and the `Signature` object directly
      assert inspect.formatargspec(*inspect.getargspec(MyClass.zzz)) == '(self, v, x=2, **kkkkwww)'

However since we are not going to drop Python2 support soon, and there is no
reffered "use-instead" functionality on Python2, let's simply silence the
warning for now.

It will be some extra work to define general equality when moving
context code into C. However that is unnecessary as usually a key is
specific to one particular package, and value associated with that key
is struct, or some other aggregate, that is holding all package-specific
context values.

Switching to compare keys by identity preserves that use-case while
making transition to C easier.

We always store just one key in there.
This will simplify transition to C.

This continues 2c8063f4 (*: Channels must be compared by ==, not by "is"
even for nilchan) and changes comparision of .done() for different
contexts from "is" to ==.

The reason is that soon we are going to move contexts into C, and this
way in context tree there potentially can be PyCtx -> CCtx -> PyCtx
chains, and while all those contexts could have the same C-level channel,
it would be difficult to propagate the knowledge that Python-level
pychan wrapper should be also exactly the same.

We already treat channel objects as pointers and require to compare them
by ==, so there should be no loss here. However to keep performance of
repeated

	select(
	    ctx.done(),
	    ...
	)

we continue to require that .done() for a particular python context
object returns exactly the same channel object every time.

Rearrange time C++ API so that e.g. time.Ticker is already
refptr<time._Ticker>. This way time users don't need to write refptr
all the time.

https://bitbucket.org/pypy/pypy/commits/251b47698e8e
https://bitbucket.org/pypy/pypy/issues/3096#comment-54707695

Which in turn fixes recover to clear current exception (see 9e6ff8bd
"golang: Fix recover to clear current exception") for context.

e.g. Context -> PyContext, with_cancel -> pywith_cancel, etc. This will
make the followup transition of context to pyx/nogil more clear.

As of 20191107 imp is present in Python 3.8 source tree and is _used_
there - see e.g.

    https://github.com/python/cpython/blob/v3.8.0-66-g7c20888e713/Lib/pkgutil.py#L188-L192

And they say imp will be there until at least Python 4 - see e.g.

    https://github.com/python/cpython/commit/e4f41deccf94

This fixes test_defer_excchain_dump failure under python3-dbg, where

E           Failed: not equal:
E           Differences (unified diff with -expected +actual):
E               @@ -1,4 +1,6 @@
E               +PYGOLANG/golang/_gopath.py:37: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses
E               +  import imp

was added into traceback printed on stderr.

Fixes: bb9a94c3 (golang: Teach defer to chain exceptions (PEP 3134) even on Python2)

Provide time.Ticker, time.Timer and friends that can be used directly
from C++ and Pyx/nogil codes. Python-level classes become small wrapper
around pyx/nogil ones.

This is the first patch that moves to Pyx/nogil classes that are
dynamically allocated on heap. refptr<T> is used to automatically manage
lifetime of such objects. At Pyx level exposed API is very similar to
Python-one, while internally it uses refptr<T> and friends.

Refobj was just added in previous patch as the class that provides base
functionality for reference-counted objects.

Since we are going to move more pygolang functionality into C++ (timers,
context, ...), and since C++ does not have garbage collector, we will
need to find a way to automatically manage memory in leak/error free way.

Let's do this via refptr<T> smart pointer (inspired by WebKit's RefPtr<T>),
which, similarly to chan<T> automatically manages pointed object's
lifetime through reference counting.

refptr<T> will be used in follow-up patches.

Top-level documentation is TODO.

This will be needed in Python-wrappers for C-level code which exposes
channels. For example in time package whose implementation will soon be
moved to nogil world fully.

3121b290 (golang: Teach pychan to work with channels of C types, not
only PyObjects) added ability to convert pychan to chan[T], but
described that ability only briefly in README. Provide more details in
that description before we add description and implementation for
reflexive pychan <- chan[T] wrapping.

We already have some sync functionality implemented in C++ (e.g.
sync.Once, sync.WaitGroup) and we are going to add more and also move
timers implementation to C++. It is getting crowded for that
functionality to still live in libgolang.{h,cpp}

-> Split sync & time functionality into their own C++ packages (still
built-into libgolang.so)

Briefly describe provided functionality instead of only referring to
https://golang.org/pkg/time.

Briefly describe provided functionality instead of only referring to
https://golang.org/pkg/sync.

Make built extensions depend on golang/{sync,time}.pxd and rebuild if
those files change. A bit overkill to depend globally, but more correct,
since an extension that is using pygolang can be using those packages.

Before the patch - without those depends - since distutils does not
implement proper dependency tracking, extensions - even those that
actually depend on sync/time pxd, were not rebuilt.

sync.pxd was missed to be added in 34b7a1f4 (golang: Expose Sema and
Mutex as public Python and Cython/nogil API). time.pxd was missed to be
added in ce8152a2 (pyx api: Provide sleep).

Just cosmetics - easier to oversee the code.

Ticker: self -> pytx
Timer:  self -> pyt

e.g. tick -> pytick, Timer -> PyTimer. etc. This will make the followup
transition of time to pyx/nogil more clear.