Let plugins make references that survive across incarnations and nodes

[akbashev]

Some actors are run by a plugin, like the cluster singleton or a virtual actor. The plugin can stop such an actor and start it again, possibly on another node. Each time, the actor gets a new ActorID.

So holding an ActorID isn't enough: once the plugin restarts the actor, calls to the old ID fail. Callers need a reference that the plugin always routes to the actor that's running now.

The singleton plugin builds such references with interceptCalls, but:

• interceptCalls is internal, so plugins outside DistributedCluster can't use it;
• the routing lives only on the node that made the reference, so a reference sent to another node stops being routed there.

What changed

All additive. Existing behaviour is unchanged.

• RemoteCallInterceptor and interceptCalls(to:metadata:interceptor:) are public. interceptCalls no longer ignores its metadata.
• forwardCall(to: ActorID, …) / forwardCallVoid deliver a call to an actor by ID, locally or remotely.
• ClusterSystemSettings.propagateMetadata(\.key) opts a metadata key into travelling with IDs. It also fixes custom metadata being dropped by ActorID.init(from:).
• ActorLifecyclePlugin.interceptor(for:) (default nil) lets a plugin intercept IDs it recognises when they're decoded on another node.

Result

cluster-virtual-actors' getActor returns such a reference. A reference kept on another node keeps working after the actor is deactivated: the next call starts the actor again and reaches it. If the actor can't be started, for example because no node hosts virtual actors, the call fails with that error.

Open questions

• Protobuf still drops custom metadata (FIXME Generalized ActorID metadata encoding/decoding #987).