Redesign actor_addr as value type - #2409

The only purpose of `actor_addr` is to select actor handles that are
stored as either `actor` or `typed_actor<...>`. However, internally it
holds a weak pointer to the actor. At this point, only for historic
reasons.

While holding a weak pointer internally works in a single-process
setting, this design breaks in a distributed system. Sending a weak
pointer is deeply problematic:
- remote nodes hold *strong* references to actors they are communicating
  with; once the remote actor terminates this reference will be dropped
  automatically from the registry
- when receiving a weak pointer to an unknown actor, there is no
  sensible thing the receiver might do other than discarding it:
  * creating a proxy would immediately expire the object again
  * there might exist weak references to the old proxy, creating a new
    one would mean those pointer would be unequal even though they
    represent the same actor
  * storing weak references in the registry would only introduce new
    problems w.r.t. proxy lifetimes, etc.

Since the purpose of `actor_addr` is to identify terminated actors in an
`exit_msg` or `down_msg`, the much cleaner, safe alternative is to
re-implement it as a simple value type holding an `actor_id` and a
`node_id`: these two IDs uniquely and unambiguously identify any actor
in a distributed CAF system.

Further, sending weak pointers is fundamentally unsafe. Hence, we also
remove the type ID for `weak_actor_ptr` and remove support for it from
the serialization API.