More safely terminate process on Windows.

We need to set the result when we detect a process is dead in the act of killing it, so that in the future we don't try to wait on it again.

Fixes #1023

Also re-enable the POSIX version of those tests. They should (and do)
pass.

Compile gevent.local.py with Cython when available.

Support using the pure python local code on Cpython too; it handles adding __new__ differently than PyPy does.

This makes things *much* faster. We had to make some small code
changes to acheive this (things like directly accessing variables
instead of redirecting through object.__getattribute__), but this
didn't impart PyPy at all (in fact it may actually be faster). New
timings compared to previous commit:

| Operation             | PyPy 2        | Python 2.7    | Python 3.6    |
|---------------------- |-------:       |-----------:   |-----------:   |
| getattr local         |   76ns        |      233ns    |      206ns    |
| setattr local         |   79ns        |      232ns    |      202ns    |
| getattr subclass      |   77ns        |      238ns    |      203ns    |
| setattr subclass      |   75ns        |      238ns    |      204ns    |
| getattr native local  |    2ns        |      139ns    |       73ns    |
| setattr native local  |    2ns        |      168ns    |       98ns    |
| getattr slowdown        	|  38x 	|     1.7x 	|     2.8x 	|
| setattr slowdown        	|  39x 	|     1.7x 	|     2.1x 	|

Speed up gevent.local

Final numbers are as follows:

| Operation            	| PyPy 2 	| Python 2.7 	| Python 3.6 	|
|----------------------	|-------:	|-----------:	|-----------:	|
| getattr local        	|   79ns 	|     1580ns 	|     1120ns 	|
| setattr local        	|   86ns 	|     1600ns 	|     1090ns 	|
| getattr subclass     	|   81ns 	|     1640ns 	|     1090ns 	|
| setattr subclass     	|   84ns 	|     1560ns 	|     1090ns 	|
| getattr native local 	|    2ns 	|      139ns 	|       73ns 	|
| setattr native local 	|    2ns 	|      168ns 	|       98ns 	|

So far, this makes accessing simple attributes about 3.5x faster on
Python 3.4 (I haven't repeated benchmarks for other platforms yet).

Refs #1020.

Fix a typo in CHANGES.rst
[skip ci]

Return the correct value from Popen.communicate() on Py3/text mode

With tests, verified against the stdlib.

Fixes a todo, and fixes #939.

Update to PyPy3 5.8.0

Fixes #1001.

Drop Python 3.3 remnants. Fixes #997.

Extract list of ignored exceptions in pywsgi to an ivar. Include errno.WSAECONNABORTED on Windows.

Extract list of ignored exceptions in pywsgi to an ivar. Include errno.WSAECONNABORTED on Windows. Fixes #999.

Specify required python versions in setup.py

Fixes #995.

Build and test with Cython 0.26