Don't immediately run links added to an Event/AsyncResult while it is running callbacks. Fixes #771.

Make the forewarn about 2.6 a permanant part of the readme that ends up on PyPi. It's not necessary to change whatsnew_1_1 in this branch, as RTD deploys from master. [skip ci]

Issue some warnings for patch_all with invalid parameters

Update socket docs to clarify their duplication of stdlib interface; tone down the warnings. [skip ci]

See changelog.rst for details.

Also skip the 'make' step on Windows if installing from an sdist (piggyback the Unix method of looking for Makefile/Makefile.ext). Ref #757.

Notes about cython on windows. XXX: Duplicated a couple places, intro.rst and README.rst, we should refactor.

- Kill older builds for the same PR automatically.
- Separate building and running tests.
  This lets us build and install a wheel up front so we don't have to do
  it again after running the tests (to distribute). This shaves 1+
  minute off each build for faster overall feedback.

Ignore the BlockingIOError name in server.py if running pyflakes on Py2; also sort the error lines by filename so they're easier to find. [skip ci]

Link socket and SSLSocket objects to their version appropriate
documentation on python.org.

Explicitly use the version-specific base socket class for SSLSocket.
This prevents the docs from listing the wrong base (because the docs are
built on Python 3 and so gevent.socket would be gevent._socket3, leading
the base of gevent._ssl2.SSLSocket to be gevent._socket3.socket).

Also, no need to 'from io import BlockingIOError' for Python 3. It
became a builtin in Python 3.3 which is the lowest version we support.

Move the Event/AsyncResult class-level _todo to a local, making classes cheaper to construct, and the _notify_links method and its interaction with _check_and_notify a bit less brain-teasing

http://bugs.python.org/issue13502

A race condition in Event that caused an improper return value if
greenlets were waiting while the event was set, but then immediately
cleared.

This was fixed in Python 3 but not (yet?) backported to Python 2.7, even
though the bug identifies that as a target.

Not only does this behaviour seem more obviously correct to me, it's
more future-proof: user share of Python 2 should be diminishing while
Python 3 share (with the correct behaviour) grows.

This required a change to our test_threading_2.py: we weren't testing
patched RLocks at all! Now we run the full set on them.

Also add test_threading.py from Python 3.4 and 3.5 to the test suite to
identify any other differences. A few tests have to be commented out
because they rely on reprs or internal class details, but nothing too
severe or concerning.

Locally tox 3.3, 3.4 and 3.5 are all green.