Instead, we only start checking the timer after the first attempted
send. This is what the C implementation in CPython does.

Fixes benoitc/gunicorn#1282.

This makes building work on platforms lacking /bin/sh.

Python 2 issue only.

Allow setting attributes on sys.modules['gevent.signal']; they go to the
gevent/signal.py module for symmetry with getattr and dir.

Reloading results in leaks so don't bother checking for them.

sslgte279 and ssl3: Move the check for _sslobj to inside the loop
instead of above it. That way if the socket is closed while
reading/writing (actually, while waiting for the read or write event)
then the ValueError will be raised on the next loop iteration (because
closing cancels the read/write events) instead of a TypeError.

This may or may not obviate the chances for an OSError: closed to be
raised.

Make _sslgte279.py preserve the create_connection function from
gevent.socket instead of from stdlib ssl.

Don't run 'test_openssl_version' when ssl is backed with LibreSSL

On LibreSSL OPENSSL_VERSION_INFO is always (2, 0, 0, 0, 0), so the test fails
for all LibreSSL versions after 2.0.

sdists built on PyPy should be the same as those built elsewhere.

Also fix the module name of gevent.libev._corecffi.

Move 3rd party libraries to deps/ directory.

Move gevent and greentest to src/ directory. This fixes tox.

At this point the builds should all work the same. Lets check.

This makes the builds faster due to smaller .c output to merge. Plus all
the improvements in Cython itself.

*NOTE* Not taking advantage of the new @property supporty yet, in case
 there are any changes to the .ppyx file we need to backport to 1.1.

Move cdef and source from Python strings to .c files that can be parsed and validated by a real C compiler.

reorganize changelog entries for 1.2a1 to group related entries. it was getting hard to follow. [skip ci]

Actually it wasn't quite as dire as described there. Instead, there were
just extra copies of each chunk of data the application yielded.

This commit removes the copies, sending the data directly to the socket.
This should be most noticeable on larger chunks.

Also re-organize the call to close() the app iter closer to where the
app iter gets assigned, making it easier to spot. Add a test for
this---turns out the validator was closing the iterator for us!

I found one genuine cycle in callback functions. The rest I can't find
any legit leak, so for now I'm disabling that part of the test on the TPE.