We need to retry if MQTT discovery fails.

[amrc-benmorrow]

Sometimes on initial deployment an MQTT client attempts to connect before the MQTT server has registered itself with the Directory. Currently this results in mqtt_client returning null, which usually results in an application crash.

We need to consider the desired behaviour here: either the client should retry, in which case the library needs to handle this, or the client should crash, in which case the library needs to do that. For some reason K8s doesn't always seem to be restarting pods that fail for this reason; possibly I'm just being impatient.