I had thought that Behaviors.setup would only get called once for an actor. This is not the case - it is called each time a behaviour is transitioned to.

There was only one place where my misunderstanding mattered, which was where a promise could be tried to be completed multiple times.

We observed an `InvalidActorNameException` when creating child actors. This could have been due to the following sequence when receiving a “Publish Received Locally” message:

1. prl received, producer actor created
2. producer actor terminates and sends a termination message
3. prl received before termination message is received, the parent actor creates another producer
4. the termination message is received and then attempts to create another producer with the same name

The solution is to explicitly track active consumers and producers rather than rely on another data structure such as `context.children`, which will be updated in response to other events.

Duplicate publish events received from a remote were previously stashed until any existing handling was complete. This commit changes that so that they are routed and handled.

The commit also tightens up some exception handling around benign exceptions that the developer should not need to consider.

In particular, actor names need to avoid the sequence number of an actor conflicting with MQTT topic names.

Packet ids were previously not distinguished according to their connection. This caused a problem when there was more than one client issuing the same packet ids!

This commit addresses the problem by permitting the routing of remotely received packet ids by connection id and client id.

Accidentally removed topic filtering from a previous commit. This could cause unexpected behaviour given that every publish would be broadcast to all topics.