[core][2/2] Kill worker on root detached actor died.

[rynewang]

PR #44214 fixes issue that when a job finished, there are workers leaked. However if a worker is assigned to a detached actor, they still leak. This PR fixes this by listening on actor death and kill assigned workers accordingly.

Based on #45637 which adds SubscribeAllActors to GcsClient. This PR uses it in node_manager and kills workers. node_manager listens to all actor changes in NodeManager::HandleActorUpdate, and it kills running workers and invoke worker_pool callback to bookkeep the died actor. worker_pool will then kill workers in periodic loops and worker start callbacks.

There are several kinds of workers to kill:

1. running workers, kept in NodeManager::leased_workers_, killed in the subscription callback in NodeManager::HandleActorUpdate
2. idle workers, killed in WorkerPool::TryKillingIdleWorkers -> WorkerPool::KillIdleWorker
3. pending pop-from-pool workers, fail in WorkerPool::PopWorkerCallbackInternal
4. pending push-to-pool workers (just started, or returned from lease), fail in WorkerPool::InvokePopWorkerCallbackForProcess

...and we treat them accordingly.

Fixes part of #45596.