Make loops automatically break when their owning hub is destroyed.

Tweak wording about Fedora package managers in docs/development/installing_from_source.rst. [skip ci]

Fixes #1689 with thanks to Fabian Affolter.

Supercedes and Fixes #1695; thanks to Anatoli Babenia.

And stop running callbacks.

This fixes #1686 and fixes #1669.

Refactor _socket2/_socket3 to move more truly common methods into SocketMixin

Move socket objects fully to __slots__.

Fixes #1724

Make gevent.pywsgi stop dealing with chunks when the connection is being upgraded

Let the application have full control over input and output.

Fixes #1712.

Per https://github.com/actions/setup-python/issues/171#issuecomment-747917974

Install thread profiling/tracing hooks in ThreadPool worker threads while the task runs

Try killing the pool to see if that buys us what we need.

Fixes #1670

Fixes #1719

Fix #1663 by more gracefully handling non-type arguments.

The documentation specifically says that kill, killall, etc, all take a a type, but
allow passing exception instances.

Also allow passing arbitrary objects; this is also not documented, and not recommended.
The result will be raising a BaseException in the greenlet. That should be better than
silently printing something to stderr, which is what happened before.

Rework the way Semaphore deals with cross-thread usage

They seem stable now.

LDFLAGS is needed on the 2010 image (for clock_gettime). The 2014
image needs -pthread (for pthread_atfork, via cffi)

Also some tweaks to the make-manylinux script; we seem to not be getting the loop I expect.

Implement timeouts for the no-hub case using spin locks.

This should ultimately be far more robust, and it handles LoopExit cases much better.

Fixes #1698

Not yet complete; On Python 2, I sometimes get ValueError: over-released the semaphore for reasons I don't understand.

Also, Python 2 in general, and PyPy in particular, sometimes fails to make progress on the semaphore tests (one of the background threads doesn't complete for some reason). The RLock tests are much worse on Python 2 also. I don't understand why. Then again, I don't understand how the normal RLock could possibly actually be thread safe either: I think it must lean heavily on the GIL.

More gracefully handle errors pickling a traceback.

Fixes #1704.

Refs https://github.com/gevent/gevent/issues/1667

Python 2, subprocess: Let unbuffered binary writes to popen.stdin loop to write all the data