Obey SDAM's requirement that if an isMaster call fails for a server that was previously available, retry it once.

I think this means that if a mongoc_async_cmd_t fails for a server that had been in a state besides MONGOC_SERVER_UNKNOWN, it should be replaced with a new mongoc_async_cmd_t, in the midst of a single call to mongoc_async_run, and it should be granted connectTimeoutMS from that moment until it times out or fails. Then the server is really unknown and shouldn't be tried again during this mongoc_async_run call.