add retry in zmqrpc

[delulu]

When I'm running Locust in distributed mode, I've noticed an exception thrown from slave agent "AssertionError: Only one greenlet can be waiting on this event".

On investigating this issue, I find out that there're multiple greenlets in SlaveLocustRunner as shown in

locust/locust/runners.py

Lines 384 to 388 in 87a9aa1

	self.greenlet.spawn(self.heartbeat).link_exception(callback=self.noop)
	self.greenlet.spawn(self.worker).link_exception(callback=self.noop)
	self.client.send(Message("client_ready", None, self.client_id))
	self.slave_state = STATE_INIT
	self.greenlet.spawn(self.stats_reporter).link_exception(callback=self.noop)

and they're talking to Locust master with the same zmq socket, so there might be concurrent calls to self.client.send, then the exception above will be thrown.

We can leverage semaphore to avoid the concurrent calls, but there will be a performance cost. Since these concurrent situations are not very likely, it's better to simply add a retry for socket communication and makes Locust more robust.

[delulu]

@cgoldberg for awareness, please have a review and let me know if any concern.