Clean up SSL exceptions and add tests for #317 and SSL timeouts in general

Additional tests for timeout/exception handling of ssl sockets; make behaviour consistent with stdlib (sendall does what python 3.5 does); fixes #317

Typo fix. Corrected spellings.

Use modern CFFI methods; drop unsupported PyPy versions < 2.6.1

No need to run fulltoxtest on PyPy---it doesn't work on Windows so the threaded file ops won't be used, and c-ares isn't a recommended configuration.

Give up on PyPy/Win for now due to the linking issues. Don't include libev_vfd if we don't need it. Update install.ps1.

Try make steps in a separate file on win and try to resolve link errors with an earlier setuptools import.

Use libev_vfd in corecffi for better Win32 compat (still not perfect). Fix a bug changing the fd and events of a io watcher and test it. Test pypy/windows.

This lets the CFFI backend be used on CPython and is more maintainable
going forward. Initial benchmarking of the CFFI backend on CPython
shows it to be a few percent slower than the cython backend, but
probably not significantly so.

Note that the tests are not expected to pass on Travis/pypy yet,
because Travis pypy is too old. Use of pyenv may be required.

Explicit test and documentation that fork-watchers and gevent.subprocess do not work in background threads. Fixes #688.

Also simplify the Timeout/if timer idiom.

Eagerly import the standard loop class when the hub is imported, preventing import lock issues under Py3. Fixes #687.

crash in a particular use cases. Fixes #237 and fixes #238.

Also clean up usage of hub._threadlocal, eliminating many try/except cases.

The particular crash seems to be in a corner-case usage (generally,
destroying a hub and/or loop seems to be advanced or rare usage), so I
feel pretty safe merging it into a release-track branch. The lifetime
management of the loop may not be totally ideal, but it solves the
issue. loop objects may now live longer (until the death of their
thread) if the hub was destroyed but the loop wasn't.

PyPy: Fix a corner case of an exception being raised by loop.handle_error during a callback that prevented other callbacks from running; this could ultimately lead to state corruption and *maybe* some crashes or leaks.

corecffi: Add a test for timer.again() and make sure timer.again() isn't subject to memory corruption that can crash the process.

Be more careful about the lifecycle management of watcher objects that may stop themselves. Fixes #676

Let GreenFileDescriptorIO inherit RawIOBase's read() method so that it handles negative/missing size reads appropriately. Fixes #675.

Always use the CountingHub, even when running leakchecks. Some tests, like test__issue607, depend on subclasses of ExpectedException not making it to the parent. The previous change to only use it sometimes (like in the past) broke those tests.