Handle the case where the coordinator is replaced with a new host

[gotascii]

If the group coordinator is replaced with a different host, but the broker id remains the same, the client will go into and endless reconnection loop. This PR refreshes the cluster data if there is a ConnectionError when joining a group. The issue is reproducible by following these steps:

• Start up a cluster with 3 nodes.
• Publish some messages to a topic.
• Connect to the topic and start an each_message loop.
• A broker, say #0 for example, becomes memoized in @coordinator in ConsumerGroup.
• Stop the each_message loop but do not exit the process.
• Kill broker 0 and bring back a new host with a different ip as broker 0.
• With the same consumer instance, run the each_message loop again.

When the above steps are taken:

• ConsumerGroup#join is called.
• Then coordinator.join_group on ConsumerGroup L:117 fails with ConnectionError.
• ConsumerGroup#join sets @coordinator = nil.
• Cluster#get_group_coordinator asks a broker for the broker id of the coordinator which is 0.
• connect_to_broker pulls cached info for id 0 (i.e. the old IP).
• Then coordinator.join_group on ConsumerGroup L:117 fails with ConnectionError restarting the loop.

Seeing as the retry for a ConnectionError is guarded by a sleep 1 I'm hoping this is a pretty safe place to refresh metadata.

[dasch]

❤️

mark_as_stale! doesn't by itself refresh the metadata, it just means that a subsequent query of metadata will cause of refresh rather than serving cached info.