Initiate Coordinator Reconnect w/ Backoff from Heartbeat Thread

[dpkp]

Attempt to fix issue raised in #2672 and possibly also #2667

If the coordinator node becomes unresponsive it can cause the consumer to busy-loop until the next attempt to reconnect to the coordinator. This happens when there is no pending request to the coordinator node itself (either a heartbeat or commit) when the broker connection fails. If the consumer is configured with auto-commits then the next commit attempt will see the coordinator failure, mark the coordinator dead, and the consumer loop will revert to waiting for the coordinator to return. But if auto-commits are disabled, the heartbeat thread will be in a limbo state where it thinks the coordinator is connecting and pauses heartbeat request. With neither a failed connection attempt or a failed heartbeat request, the coordinator is never marked dead and the consumer busy loops. To fix, this patch adds an explicit connection attempt to the heartbeat thread loop whenever the coordinator is known but disconnected. If the connection fails the client will mark the node as disconnected and add a backoff reconnect delay. The non-zero backoff delay causes the consumer to mark the coordinator dead, which prevents the busy loop.