More attempts to address heartbeat timing issues in consumers, especially with the iterator interface. Here we can reduce the `client.poll` timeout to at most the retry backoff (typically 100ms) so that the consumer iterator interface doesn't block for longer than the heartbeat timeout.

This splits the `_find_coordinator_id()` method (which is blocking) into
request generation / response parsing methods.

The public API does not change. However, this allows power users who are
willing to deal with risk of private methods changing under their feet
to decouple generating the message futures from processing their
responses. In other words, you can use these to fire a bunch of requests
at once and delay processing the responses until all requests are fired.

This is modeled on the work done in #1845.

Additionally, I removed the code that tried to leverage the error
checking from `cluster.add_group_coordinator()`. That code had changed
in #1822, removing most of the error checking... so it no longer adds
any value, but instead merely increases complexity and coupling.