This change fixes a race conditon related to _WaitHandleFuture.cancel() leading
to Python crash or "GetQueuedCompletionStatus() returned an unexpected event"
logs. Before, the overlapped object was destroyed too early, it was possible
that the wait completed whereas the overlapped object was already destroyed.
Sometimes, a different overlapped was allocated at the same address, leading to
unexpected completition.

_WaitHandleFuture.cancel() now waits until the wait is cancelled to clear its
reference to the overlapped object. To wait until the cancellation is done,
UnregisterWaitEx() is used with an event instead of UnregisterWait().

To wait for this event, a new _WaitCancelFuture class was added. It's a
simplified version of _WaitCancelFuture. For example, its cancel() method calls
UnregisterWait(), not UnregisterWaitEx(). _WaitCancelFuture should not be
cancelled.

The overlapped object is kept alive in _WaitHandleFuture until the wait is
unregistered.

Other changes:

* Add _overlapped.UnregisterWaitEx()
* Remove fast-path in IocpProactor.wait_for_handle() to immediatly set the
  result if the wait already completed. I'm not sure that it's safe to
  call immediatly UnregisterWaitEx() before the completion was signaled.
* Add IocpProactor._unregistered() to forget an overlapped which may never be
  signaled, but may be signaled for the next loop iteration. It avoids to
  block forever IocpProactor.close() if a wait was cancelled, and it may also
  avoid some "... unexpected event ..." warnings.