- 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.

It's a pet peeve of mine when you can't at least directly get to the
changelog for a release from the PyPI page. Before this, it took at
least three clicks and some reading to figure out where they
were.

(Including directly in the page is better, but our notes are kind
of long; maybe we should split them up?)

[skip ci]

my one-off scripts I've been using to build releases so they don't get lost. needs much love to be useful to anyone else. [skip ci]

pypy 'incorrectly' raises a plain ssl.SSLError where CPython would raise a ssl.SSLEOFError, even though pypy defines that class. But pypy doesn't try too hard to match exact exceptions with cpython.

Add unit tests for sending empty data using sendall, sendall with a
timeout, and send. Note that SSLSocket.send(b'') *does* raise EOFError,
just like the standard library.

This avoids a printed/ignored KeyError seen on PyPy/arm. [skip ci]

This showed we were missing implementations for socket.sendfile and
get/set_inheritable.