Resuming a paused actor detaches it when its node has no room - #2598

A PAUSED actor's snapshot sits on the local disk of the node it paused on, so it could only resume there. If that node was full, the resume failed, even if other nodes had free workers. Now the resume runs these steps:

1. Try to get a free worker on the actor's own node. If there is one, restore there from the local snapshot, as before.
2. If the node is full, check whether any other node has a free worker. If none does, stop: the actor stays PAUSED and the resume returns ResourceExhausted, so the caller can retry later.
3. Mark the actor SUSPENDING, and record where in object storage its snapshot will be uploaded. That record is marked PENDING until the upload finishes.
4. Ask atelet on the old node to promote the snapshot: upload it to object storage, then delete the local copy.
5. Mark the upload COMPLETED, forget the old node, and mark the actor SUSPENDED, all in one update. Then delete the uploaded snapshot this one replaced, if any. Failing that delete does not fail the resume.
6. Get a free worker on any node, restore the actor there from the uploaded snapshot, and mark it RUNNING.

Steps 3 to 5 are the same steps SuspendActor runs for a PAUSED actor, so the snapshot is uploaded with the template's preferredFidelity, as a suspend would do.

If the upload in step 4 fails:

• If atelet is unreachable, the actor stays SUSPENDING and the resume returns Unavailable. Its local snapshot is still on the old node.
• For any other error, the actor is crashed, the same as when SuspendActor fails.

Resume only starts from PAUSED or SUSPENDED, and also accepts RESUMING, to finish a resume that was cut off. A RUNNING actor returns right away. Resume refuses a SUSPENDING actor with FailedPrecondition, including one left behind by a failed step 4. To recover it, call SuspendActor, which finishes the upload to the location recorded in step 3, then resume the actor.

sequenceDiagram
    participant C as client
    participant A as ate-api-server
    participant S as store
    participant L1 as atelet on node1
    participant O as object storage
    participant L2 as atelet on node2

    C->>A: ResumeActor
    A->>S: assign worker, required node = node1
    S-->>A: no capacity (errNoWorkerRoom)
    A->>S: Schedule with no node constraint
    S-->>A: some node has room
    A->>S: UpdateActor: SUSPENDING, local snapshot durable = PENDING at snapshot URI
    A->>L1: PromoteSnapshot(snapshot uuid, fidelity, object URI)
    L1->>O: upload snapshot files, then manifest
    L1->>L1: delete the actor's local snapshots
    L1-->>A: ok
    A->>S: UpdateActor: durable COMPLETED, locality and assigned_node cleared, SUSPENDED
    A->>O: release the previous durable snapshot (best effort)
    A->>S: assign worker, any node
    S-->>A: worker on node2
    A->>L2: Restore(same snapshot)
    L2->>O: fetch manifest, download files
    L2-->>A: ok
    A->>S: UpdateActor: RUNNING
    A-->>C: ResumeActorResponse

• Tests pass
• Appropriate changes to documentation are included in the PR