While working on on a SIGINT signal handler for the ProactorEventLoop on 
Windows (issue #191), I hit a bug on Windows: _WaitHandleFuture.cancel() crash 
if the wait event was already unregistered by finish_wait_for_handle().

Work-in-progress patch to fix _WaitHandleFuture.cancel():
http://codereview.appspot.com/114250043

The patch logs also OSError in _OverlappedFuture.cancel(), I didn't test this 
change.

I also added a FIXME "why is ERROR_IO_PENDING ignored?" on 
_overlapped.UnregisterWait(), because I don't understand why the error is 
ignored. I should read the doc, but I didn't have the time for that yet:
http://msdn.microsoft.com/en-us/library/windows/desktop/ms686870%28v=vs.85%29.as
px