The current client attempts to bootstrap once during initialization, but if it fails there is no second attempt and the client will be inoperable. This can happen, for example, if an entire cluster is down at the time a long-running client starts execution.

This commit attempts to fix this by removing the synchronous bootstrapping from `KafkaClient` init, and instead merges bootstrap metadata with the cluster metadata. The Java client uses a similar approach. This allows us to continue falling back to bootstrap data when necessary throughout the life of a long-running consumer or producer.

Fix #1670

* Add BrokerConnection.send_pending_requests to support async network sends
* Send network requests during KafkaClient.poll() rather than in KafkaClient.send()
* Dont acquire lock during KafkaClient.send if node is connected / ready
* Move all network connection IO into KafkaClient.poll()

In an effort to reduce the surface area of lock coordination, and thereby hopefully reduce lock contention, I think we can remove locking from the read-only KafkaClient methods: connected, is_disconnected, in_flight_request_count, and least_loaded_node . Given that the read data could change after the lock is released but before the caller uses it, the value of acquiring a lock here does not seem high to me.

Fix #1633