Mark test__fileobject.py flaky on appveyor due to random failures that are apparently timing related.

The cert for svn.python.org has changed breaking many test_ssl.py cases.

The upstream resolution is being discussed in
https://bugs.python.org/issue25940, but the interim solution was to
disable the failing tests to avoid blocking releases. I'm applying the
same approach here for a temporary fix.

Optimize CFFI backend to use less memory by taking advantage of the libev 'data' pointer. Also simplifies the callback code.

Use a manually locked pure-Python Semaphore implementation on PyPy

Based on http://fxr.watson.org/fxr/source/sys/_types.h#L52 freebsd also
uses short for st_nlink in struct stat.

Not mentioning in the change notes because I have no way to actually
verify. But it shouldn't hurt.

More tests for various uses of locks in a trace function. Make most of them work or at least not deadlock.

The main stated reason for untraceable locks was a bug in Python 2.5,
fixed in all recent versions, so that shouldn't be a concern.

The semaphore takes some careful lock management to come close to
duplicating the thread-safety inherent in Cython's compiled methods.

'pypy -m timeit -s "import g.l.Semaphore; s=Semaphore()"' 's.acquire();s.releas()':

With Cython compiled semaphore: 5usec
With pure-python sempahore: 0.004usec
With pure-python-but-locked semaphore: 0.178usec
Cython on CPython: 0.141usec

Total test runtime is unchanged.

Use the new ssl module on any python version that natively supports it

Make microbenchmarks run on recent versions; allow choosing alternate pythons for easy comparison (E.g., using .tox/py[py]/bin/python.) [skip ci]

Exclude build products of _corecffi_build.py from the sdist, they can
cause a failure to link if the platforms don't match.

Fix a lurking GC-related crash with stat watchers on PyPy, detected by CPython re-using the memory much faster.

Make the CFFI backend handle the watcher's args property more like the Cython backend. Fixes test__core_watcher.

This allows building, using and testing corecffi on CPython, which can
make debugging easier. It also simplifies the build steps for PyPy.

Start testing corecffi on CPython too for comparison with PyPy and speed
comparisons with CPython. Currently getting a few unexpected failures
that need investigation; could be bugs in corecffi.

Remove some FLAKY tests from appveyor that should no longer be FLAKY; add test_selectors for 3.5/3.5 because it has timeout-related code that can be flaky there. Remove obsolete failure from PYPY case because we don't support that old version anymore.

test_hub_join_timeout was still flakey on Py 3.4/64-bit appveyor

One finer-grained skip for appveyor

AppVeyor: Make a number of timing-related tests less strict, hopefully fixing most of the appveyor failures

issues_671: Make more robust by sending a second signal and printing diagnostics.
Skip the few test cases that are still timing out unreliably on appveyor.
More explicit tests for gevent.Timeout functionality plus timing fixes.

Normalize recent GNU ls output to match the expected BSD-style output present in the docstrings of subprocess.py under Python 3. Fixes #681.