Don't let destroying the default C loop from two different Python loo…

Don't let destroying the default C loop from two different Python loop objects crash the process. Fixes #1098

Take a whack at using ccache to speed up builds on CI.

[skip ci]

Centralize configuration in a ``gevent.config`` object.

Extra tests for _config. Fix a long-standing bug where package paths in import names did not clean up after themselves.

Before I realized that was the root cause, I refactored to avoid
starting the hub greenlet *at all*, and that also fixed the problem.

The problem was that some versions of CPython using the libev-cext
loop (and only that one) would crash garbage collecting a
greenlet (presumably the hub greenlet, which was attached to a
destroyed loop):

0   libsystem_kernel.dylib        	0x00007fff66455e3e __pthread_kill + 10
1   libsystem_pthread.dylib       	0x00007fff66594150 pthread_kill + 333
2   libsystem_c.dylib             	0x00007fff663648fe raise + 26
3   libsystem_platform.dylib      	0x00007fff66587f5a _sigtramp + 26
4   ???                           	000000000000000000 0 + 0
5   greenlet.cpython-35m-darwin.so	0x000000010d76b297 green_clear + 183
6   org.python.python             	0x000000010c707e0c collect + 1824
7   org.python.python             	0x000000010c7076e3 _PyGC_CollectNoFail + 43
8   org.python.python             	0x000000010c6e81d4 PyImport_Cleanup + 968
9   org.python.python             	0x000000010c6f1b49 Py_Finalize + 92
10  org.python.python             	0x000000010c6f20d5 Py_Exit + 13
11  org.python.python             	0x000000010c6f4631 handle_system_exit + 320
12  org.python.python             	0x000000010c6f42a1 PyErr_PrintEx + 42

We should take a look at that and see if we can do something more graceful.

The gevent.select.poll object must keep a tighter handle on the lifetime of io watchers it spawns. Previously they were leaking if not unregistered, which some callers fail to do --- looking at you, dnspython.

This lets us document things more clearly, and, with care, allows
configuring gevent from Python code without any environment varibles.

All the fileobject implementations needed to be moved out of the
``gevent.fileobject`` module so as not to introduce import cycles that
made configuring from code impossible.

Fixes #1090

Add support for /etc/hosts to dnspython resolver.

Based on code from eventlet, but heavily refactored and modified:

- Split host parsing into its own class, re-used from the test case
- Hosts handle PTR lookups for ipv4
- Some bug fixes around what should be strings where.
- Allow strings for the rdtype paramater
- Watch last modified time instead of reloading the file at fixed
  intervals.

More of our DNS tests pass now (notable TestEtcHosts is now enabled
for this resolver), and we tend to generate more conformant results
than the ares resolver (by observation, not tested).

Let dnspython resolver be used without monkey-patching

Install openssl on travis for 3.7b1 as a fix for  https://github.com/travis-ci/travis-ci/issues/9069

[skip appveyor]

The build cache was the only thing keeping 3.7.0a4 around; it can no
longer be installed from pyenv. And I accidentaly deleted the master
build cache, so we have no choice.

It may be difficult to get 3.7b1 installed due to
https://github.com/travis-ci/travis-ci/issues/9069

Do this by importing dns specially with sys.modules containing the
gevent modules. This is the technique eventlet uses (but we don't
leave the patched modules sitting around in sys.modules after we're
done; eventlet may have good reason to do that, but this is not a
public API here and is specialized just for dns).

Cleanups for timeout.py docs.

Deal gracefully with passing None watchers to hub.cancel_wait, something that may be more common now that we try to destroy watchers ASAP for GC reasons. Fixes #1089.

Add a DNSpython resolver