I'll need help to reproduce them.

But temporarily add some extra debugging to the makefile to see if we
get a clue.

The cpython failure in test_subprocess.py definitely began with this
branch (which doesn't make much sense). The PyPy failure was happening
long before.

Skip that one test that fails on Python 3.7 under libuv on Travis. It doesn't fail for any other CPython version there and it doesn't fail locally. Chalking it up to a alpha issue for now.

Instead of manually having it schedule callbacks, we instead run
callbacks at a time in the event loop cycle equivalent to when libev
runs them. We do this with a check watcher (instead of a 0 duration
timer).

Fixes #1058

The good news is this probably explains a lot of the timer oddities we
see under libuv (see detailed comments in the code). The bad news is
it means we need to rethink some things for libuv.

Test all backends that are supported with the full test suite in a
single run. We have a high setup overhead, so not only does this
reduce the number of jobs, it hopefully reduces the overall time
taken, while actually increasing the number of tests run.

Support for Python 3.7

No new unexpected failures.

Rename the 'async' methods and classes to 'async_'. I couldn't think
of a better name. But keep the old names available for BWC. Fixes #1047.

Request more specifics.

Explicitly close IO watchers 

Disable PyPy+libuv on Windows again, still some very strange errors there. But PyPy+libuv on darwin/posix works, and CPython3.6 Windows works. Perhaps the issue is the 32-bit pypy on windows and the way it adjusts socket.fileno()?

Hopefully solve the test__os issue. It turned out to be that calling uv_close at random GC times is a very bad idea under libuv. See the comment. Try enabling pypy+libuv on posix and windows; it's working on darwin for me.

By closing them when we're done, we can remove reliance on GC to clean
up multiplexed watchers and solve problems for PyPy and Windows (and
probably eventually simplify the handle cleanup we were doing.)

This makes many more tests on PyPy pass on Darwin; down from a score
of failures to about 4, but a baffling 4. test__os:TestOS_nb times
out; investigation shows that both the read and write sides of the
pipe are waiting on events, both are in libuv, and the write side
should be writable, but it's not getting an event. Not clear why this
is. I also see it with CPython 3.6, and I don't think I did before.

Support libuv under CPython 3.6 on Windows

First the crash report:

```
  testInterruptedRecvmsgTimeout (__main__.InterruptedRecvTimeoutTest) ... Traceback (most recent call last):
    File "/home/travis/build/gevent/gevent/src/gevent/_ffi/loop.py", line 83, in python_callback
      def python_callback(self, handle, revents):
    File "/home/travis/build/gevent/gevent/src/greentest/3.6/test_socket.py", line 3698, in <lambda>
      lambda signum, frame: 1 / 0)
  ZeroDivisionError: division by zero
  Fri Dec 15 16:09:36 2017
  ok
  testInterruptedSendTimeout (__main__.InterruptedSendTimeoutTest) ... Fatal Python error: ffi.from_handle() detected that the address passed points to garbage. If it is really the result of ffi.new_handle(), then the Python object has already been garbage collected

  Thread 0x00002b26d4c01700 (most recent call first):
    File "/home/travis/build/gevent/gevent/src/gevent/_threading.py", line 152 in wait
    File "/home/travis/build/gevent/gevent/src/gevent/_threading.py", line 436 in get
    File "/home/travis/build/gevent/gevent/src/gevent/threadpool.py", line 200 in _worker

  Thread 0x00002b26d4a00700 (most recent call first):
    File "/home/travis/build/gevent/gevent/src/gevent/_threading.py", line 152 in wait
    File "/home/travis/build/gevent/gevent/src/gevent/_threading.py", line 436 in get
    File "/home/travis/build/gevent/gevent/src/gevent/threadpool.py", line 200 in _worker

  Current thread 0x00002b26cc7b2a00 (most recent call first):
    File "/home/travis/build/gevent/gevent/src/gevent/_ffi/loop.py", line 178 in python_stop
    File "/home/travis/build/gevent/gevent/src/gevent/libuv/loop.py", line 289 in run
    File "/home/travis/build/gevent/gevent/src/gevent/hub.py", line 688 in run
```

The main problem turned out to be the way in which libuv reports
signals at random times, much worse than the way libev does. This led
to callback functions unexpectedly returning 0, because that's the
default for an unhandled exception or an onerror handler that returns
None. That in turn led to calls to python_stop that we weren't
expecting using a handle that had already been deallocated.

The fix is to both be defensive about what handles we try to use, and
to be more explicit about our error returns.

Disable pypy on appveyor again because it takes too long. I need to make sure it runs on POSIX first.

I solved the test_ftplib.py windows failures locally by more careful handling of lifetimes of io watchers: they should be disposed when the fileno is. I investigated a more drastic approach using new watchers entirely every time we restart, but it doesn't seem necessary, at least on CPython (I can't even run PyPy+libuv locally on darwin though, so there are more problems lurking there).

All tests are passing locally after fixing a GC-cycle related issue in
the io watcher that kept the wrong socket watchers alive. This should
probably be generalized to all watchers on all FFI platforms.

Attempt to fix (or workaround) test__socket.py on Windows

It runs locally for me on Win10/Py3.6 with both libev and libuv.

It had been sometimes failing since 2016-02 for 2.7 and 2016-04 for
3.5.

Here's the comment in the code indicating what I tracked down. Note
that the changes are more of a workaround than a true fix:

XXX: On Windows (at least with libev), if we have a cleanup/tearDown
method that does 'del self.listener' AND we haven't sometime
previously closed the listener (while the test body was executing) we
tend to sometimes see hangs when tests run in succession; notably
test_empty_send followed by test_makefile produces a hang in
test_makefile when it tries to read from the client_file, because the
accept() call in accept_once has not yet returned a new socket to
write to.

The cause *seems* to be that the listener socket in both tests gets
the same fileno(); or, at least, if we don't del the listener object,
we get a different fileno, and that scenario works.

Perhaps our logic is wrong in libev_vfd in the way we use
_open_osfhandle and determine we can close it?

Additional note:

Possibly libuv doesn't suffer from this since it takes the SOCKET
object directly? I didn't fully test that situation, but libuv tests
were failing and are now passing.