Retry read on master when causal_reads timeout is exceeded

If a read done with causal_reads=true is done on a slave and it fails due to a timeout, the query could be retried on the master and we'd be guaranteed to get up-to-date data.

Attached a patch that implements this to the issue.