Refactor disable requests query to async + RC actor lifecycle handling

[Andyz26]

Summary

• Move loadAllDisableTaskExecutorsRequests from blocking preStart() to async pipe on default-blocking-io-dispatcher with retry (up to 5 attempts, 5s delay). A
  data store timeout during this synchronous call caused ActorInitializationException, which MantisActorSupervisorStrategy handles with permanent stop() — silently
  killing the ExecutorStateManagerActor with no recovery path.
• Add child actor death detection in ResourceClusterActor — watch children, handle Terminated by nulling stale refs (activating existing null-check error paths) and
  scheduling delayed re-creation.
• Add Terminated handler in ResourceClustersManagerActor — remove stale map entry and stop the surviving sibling actor so the full pair is cleanly recreated on next
  request.
• Add observability metrics (executorStateManagerInitFailure, childActorTerminated, childActorRecreated) with actorType tags.

Root Cause

ExecutorStateManagerActor.preStart() made a synchronous blocking gRPC call to DGW via mantisJobStore.loadAllDisableTaskExecutorsRequests(). When actual data store exceeded its
SLO:

1. CompletionException: DEADLINE_EXCEEDED thrown in preStart()
2. Akka wraps as ActorInitializationException
3. MantisActorSupervisorStrategy → SupervisorStrategy.stop() (permanent, no restart)
4. ResourceClusterActor holds dead ActorRef — all forwarded messages go to dead letters silently
5. No watch/Terminated handling at any level — entire resource cluster becomes non-functional with no detection or recovery

[github-actions]

Test Results

774 tests  +3   763 ✅ +3   10m 12s ⏱️ +2s
162 suites +1    11 💤 ±0 
162 files   +1     0 ❌ ±0 

Results for commit 70120bd. ± Comparison against base commit 406e1d7.

♻️ This comment has been updated with latest results.