[Placement group] Refine Remove API

[rkooo567]

Describe your feature request

When workers are killed, their death information is not properly propagated to the cluster. This causes issues like

• When tasks or actors are killed by pg fate sharing, their exception type is "WorkerCrashedError" or "ActorDiedError" instead of PlacementGroupError.
• When we schedule a new task on removed placement group, they become permanently infeasible. This has issues when task retry is involved (e.g., Task has failed because the worker is killed, and the core worker try rescheduling a task, but it becomes infeasible).
• Newly submitted infeasible tasks demand will never be deleted.

The ideal final state:

• All killed workers by "remove_placement_group" should raise PlacementGroupRemovedException.
• All subsequent actor creation & task submission with the placement group should raise PlacementGroupRemovedException.

Implementation:
There are 2 possible implementations

• When the placement group is removed, publish the removal information to all interest party (core worker & raylet). This means when we start using the pg, we should always subscribe the pg state (to make sure we don't use removed placement groups).
• Mark tasks as "infeasible" if the corresponding placement group is already removed. This can be implemented as a part of Raise an Exception when a task, actor, or placement group is permanently infeasible #18835

Prefer the second solution

[rkooo567]

Not worth doing it before new scheduler is enabled.

[rkooo567]

@oliverhu If we implement pubsub stuff, this can also be handled.