cython ignores the preprocessor directives we previously had in
libev.pxd, so that was doing nothing, and it was treating vfd_socket_t
just as something it had to turn into an int.

The function it used to do this was the same function it used to turn
a generic PyObject into an int, once all the typedefs were substituted
in (which does use the c preprocessor from libev_vfd.h). So it should
be functionally equivalent.

The vfd_* functions are no-op macros on POSIX, so there shouldn't be a
performance difference. The minor exception is that __dealloc__ is
always defined, and although it's a static function suitable to
inline, there is an extra Python API call to PyErr_Fetch and
PyErr_Restore. But if that makes a difference, I didn't spot it.

Move to the safer __cinit__ for all watchers.

Now that the base watcher has access to its libev ev_watcher pointer, start moving the properties there, starting with ref.

We no longer need to gc before getting a libuv io object on PyPy: everything is deterministically closed (or should be).

Debugging for a weird EPERM error in test__fileobject.py on travis with libuv; not seen locally or on appveyor [skip appveyor].

And must be close() when done. The IO watchers already required this
for libuv; doing it for everything simplifies memory management and
object lifetimes and lets us get rid of a bunch of weakref objects
with callbacks.

This is a new requirement and will only affect those that use
CFFI (e.g., PyPy) and specifically it's needed for libuv; libev
doesn't really do much with it.

Some debugging help for forgetting to close objects piggybacks on
tracemalloc on Python 3.

test_ftplib.py fails in Python 3 because of some warnings about
unclosed objects.
test__backdoor.py fails in Python 2 for the same reason.

[skip appveyor]

GEVENTSETUP_EV_VERIFY has to be set when gevent is compiled. Move it from the 'leaktest' target down to where we call develop