In the case we're using a threadpool, depending on where the threads are waiting in completion queue's next, they'll get stuck on shutdown because we're only kicking one of them. Instead, let's cascade-kick the pollsets when we shutdown so that we make sure all of them are exitting properly.

Fixes #2190. In grpc_pollset_work we need to pass down which lock to unlock before actually doing something.

Fixes https://github.com/grpc/grpc/issues/2194 .

Part of fixing https://github.com/grpc/grpc/issues/2192

Part of fixing https://github.com/grpc/grpc/issues/2192

Part of fixing https://github.com/grpc/grpc/issues/2192

Part of fixing https://github.com/grpc/grpc/issues/2192

- grpc ruby 0.9.3 should have released from the release branch but was not
- it needs to be re-released from that branch

It shouldn't count towards coverage

Shutting down is synchronous under Windows, so we don't want to do that while holding a lock.

With the backup poller gone, the caller of grpc_pollset_work is expecting it to sleep. Aligning the behavior of the Windows port with the behavior of the Posix port.

Since alarm checks may mutate work deadlines for pollsets, the value
returned from maybe_work is meaningless. Instead, maybe pollset_work
always return 1 if maybe_work is invoked, and then redo the deadline
check _on the next call_ to pollset_work.

And free ops_array that was leaking.

- introduces a #wait method on the call operation view
- invokes #wait on a Notifier that is created for all operations
- ensures the Notifier is invoked if necessary whenever a client request
  completes
- updates the interop_test to use op.wait before checking if the call
  was cancelled.