[ci skip]

we no longer stores test results in a [sqlite] database

The stdlib tests are now included in the repository. since ubuntu
does not ships them in any package.

The testrunner is completely rewritten. When testrunner.py was originally written
a lot of tests were failing and some of them were hanging. Thus testrunner.py
was focused on proving good reports and could retry the tests with a particular
test case disabled if it detected time out. Now testrunner is somewhat simpler as
we mostly concerned whether the tests pass or not. It also runs the tests
concurrently, to improve the speed at which we get the results back from travis-ci.
Some tests time execution and assert certain timings, this can fail, especially when
the number of workers is bigger than cpu count. The new testrunner will re-try
tests that failed in a sequential run, when nothing else is run. If this fixes
the problem, the test is reported as "succeeded, but only on the second try" and does
reports the whole suite as failed.

.travis.yml is fixed to download binary packages for all dependencies and use system python.

rename test__benchmarks.py to xtest__benchmarks.py to exclude it form testrunner.py.

Exclude test__count from test_thread.py, because it uses function thread._count() we don't implement

Run monkey patched tests twice, first with patch_all(), then with patch_all(Event=True)

if True, threading.Event is replaced with gevent.event.Event

Previously this was handled by 'callback' setter,
but that one now accepts both callable and None, so
we need a separate check for None in start()

this prevents test__all__.py from failing and prevents monkey from patching it

(at least on linux)

that way it does not fail when being used as class member

it's always the opposite of _WIN32
this speeds up cythonpp.py part of the build

For non-network requests, like "localhost", it seems to be quite hard
to make gevent.ares to do the right (=compatible with stdlib) thing.

So we check if host does not have '.' and if that's the case,
we fallback to using threadpool.

This covers requests like "localhost", "ip6-localhost" and Python-specific
things, like "<broadcast>"

It's possible to disable this by passing use_threadpool=False to ResolverAres
or by settings resolver.pool = None.

For string ports (e.g. "http") resolver_ares/getaddrinfo previously only
consulted getservbyname(port, "tcp") or getservbyname(port, "udp"), but never both.

It seems that checking of both work (and creating result based on that) makes for
better compatibility with stock getaddrinfo, at least on linux.

it is not handled by test__monkey_patching.py

that way test runner does not fail on ubuntu where this tests are not available
a new test__monkey_patching.py is added which runs all stdlib tests

it can mask legitimate EINVAL cases, causing busy loop

this reopens issue #148

core: allow 'callback' property of a watcher to be set to None; del .callback is no longer supported

This fixes links execution order to be the same as they were added, fixing issue #143
The microbenchmarks show that it's not slower, in fact, a bit faster.

subprocess: add 'timeout' parameter to wait() in Windows case to avoid raising TypeError (timeout is ignored on Windows for now)