Add a fail-back polling mode for Windows+3.8+

src/python/grpcio/grpc/_cython/_cygrpc/aio/completion_queue.pyx.pxi

@@ -17,6 +17,8 @@ import socket
 cdef gpr_timespec _GPR_INF_FUTURE = gpr_inf_future(GPR_CLOCK_REALTIME)
 cdef float _POLL_AWAKE_INTERVAL_S = 0.2
 
+cdef bint _no_fd_monitoring = False
+
 
 IF UNAME_SYSNAME == "Windows":
     cdef void _unified_socket_write(int fd) nogil:
@@ -47,7 +49,14 @@ cdef class _BoundEventLoop:
             handler,
             loop
         )
-        self.loop.add_reader(self.read_socket, reader_function)
+        # NOTE(lidiz) There isn't a way to cleanly pre-check if fd monitoring
+        # support is available or not. Checking the event loop policy is not
+        # good enough. The application can has its own loop implementation, or
+        # uses different types of event loops (e.g., 1 Proactor, 3 Selectors).
+        try:
+            self.loop.add_reader(self.read_socket, reader_function)
+        except NotImplementedError:
+            _no_fd_monitoring = True
 
     def close(self):
         if self.loop:
@@ -97,7 +106,15 @@ cdef class PollerCompletionQueue(BaseCompletionQueue):
                 self._queue_mutex.lock()
                 self._queue.push(event)
                 self._queue_mutex.unlock()
-                _unified_socket_write(self._write_fd)
+                if not _no_fd_monitoring:
+                    _unified_socket_write(self._write_fd)
+                else:
+                    with gil:
+                        # Event loops can be paused or killed at any time. The
+                        # most robust way to make sure someone is polling is
+                        # awaking all loops up.
+                        for loop in self._loops:
+                            self._handle_events(loop)
 
     def _poll_wrapper(self):
         with nogil: