Some errors are fatal, so we should log the exception.
Other errors we will try to handle and try again, in such a case
just log a warning unless it is the final attempt. This means
successful retries will not result in an exception being logged.

- SlotNotCoveredError, was re-raising the exception one attempt too soon.
- Similarly log_exception

populate_startup_nodes is meant to put all the known nodes in the startup nodes list,
but when it was called, the nodes list it was using (self.nodes) is not populated yet.
So populate startup nodes after initialising self.nodes,
otherwise when we choose a random connection (which is chosen from startup nodes),
the options are too limited. Often a user will specify a single startup node at the beginning.

Add a docker-compose.yml file configuring a redis cluster instance for running the tests against.