Auto-track and globally shut down all Forge actors and services

[DNXie]

Context: #360

This PR adds automatic allocation tracking for all ForgeActor and ServiceInterface instances via the global Provisioner.
Actors and services are now automatically registered with the provisioner when spawned through .as_actor() or .as_service(), enabling unified lifecycle management and eliminating the need for manual shutdown calls.

Changes

• ForgeActor.as_service()
  Registers the top-level ServiceInterface proxy (instead of raw ActorMesh) after initialization.
• ForgeActor.as_actor()
  Registers the spawned ForgeActor proxy after successful setup.
• Provisioner.shutdown_all_allocations()
  Handles both ForgeActor and ServiceInterface cleanly; internal ActorMesh handles are no longer registered.
• Improved logging for spawn and teardown events.

Result

All actors and services are now automatically tracked and can be shut down gracefully with a single call.

Before

await policy.shutdown()
await ReplayBuffer.shutdown(...)
...
await shutdown()

After

await shutdown()

Test:

python -m tests.sandbox.toy_rl.sumdigits --config tests/sandbox/toy_rl/sumdigits.yaml

The log after ctrl+C (cleaned version):

...
loss/training_step: -0.6666666865348816 at training step 1
^CShutting down...
Shutting down actor DatasetActor
Shutting down service Policy
Shutting down actor Trainer
Shutting down actor ReplayBuffer
Shutting down service RewardActor
Shutting down service RefModel

TODO:

Merge

await mlogger.shutdown.call_one()

in the centralized shutdown.

Will do it after the Monarch error with the logger (#360 (3)) is fixed. Fixing it is out of scope of this PR.

[DNXie]

Adding this quiet check because otherwise when we shutdown a service, it would call actor.shutdown and print the log twice.

[allenwang28]

hmmm, I think an even simpler approach is to just track the proc meshes right? We can just do await proc_mesh.stop() and I think everything inside of it should shut down neatly. Let me know if that doesn't work though

[DNXie]

I think it would work for actor since shutting down actor is essentially stopping the proc_mesh.
But for service, it involves some other operations such as stopping the replicas and healthy loop.