Stop keep alives when worker reconnecting to the scheduler

[jacobtomlinson]

Related to #3488.

I don't think this fixes the problem, but I think it should address the CommClosedError exceptions which are seen in the log.

When a worker registers with the scheduler it starts sending keepalive messages via a periodic callback. However that never seems to stop if the connection is broken. If the connection hangs for a long period of time the worker still attempts to send the keepalive messages.

This PR moves the definitions of the BatchedSend and keep-alive callback to the __init__ and stops the callback during the reconnect. This is already being done for the heartbeat.

[jacobtomlinson]

I'm not entirely sure what I've broken here. Assistance would be appreciated.

[mrocklin]

I'll try to take a look later today

[mrocklin]

If you run tests with -s you'll see errors in the logs that look like the following

>       self.batched_stream = BatchedSend(interval="2ms", loop=self.loop)
E       AttributeError: 'Worker' object has no attribute 'loop'

I've pushed a patch moving the construction a bit lower, but on my machine there are still some failures. It might make sense to put this construction back in `start`` if there isn't a big reason to move it out.

[jacobtomlinson]

Thanks for pushing to this @mrocklin. The CI seems happy here.