Obey Server Discovery and Monitoring Spec's rules for parsing and responding to "not master" errors: mark the server Unknown as if it were a network error. Detect these errors in the topology scanner's async command callback and in mongoc_cluster_run_command_monitored / mongoc_cluster_run_command_internal.

This has been a missing feature since we implemented SDAM in libmongoc
1.2.0. It's now urgent to fix because Retryable Writes depends on this
feature. Otherwise during a stepdown, the driver might try a write
against a former primary, get a "not master" error, and try again with
the same server. Marking it Unknown ensures we re-scan the topology.