anytrait: evaluate performance consequences of using per-trait listeners instead of per-object listeners

The filtered trait mechanisms in the `observe` framework (and in particular the `anytrait` special case introduced in #1496) register listeners on a per-trait basis.

This is in contrast to the `on_trait_change` "anytrait" handling, which registers listeners on a per-object basis.

It would be worth analysing the consequences of this choice on performance and deciding whether there's a performance case for looking at migrating the filtered trait observers to use the object-based notification.