Filter stale worker addresses during cleanup via scheduler-authoritative data

[garvit-gupta]

Context

Follow-up to #1045. Cleanup paths (Recovering, Failing, Stopping) can iterate stale gRPC channels targeting addresses recycled to other workers. While #1045 prevents misrouted messages at the receiver, this change prevents them at the sender by filtering the in-memory worker map against the scheduler's authoritative view before sending.

Changes

• Add rpc_address tracking to WorkerStatus so cached channels can be matched against live addresses.
• Add worker_addresses_for_job to the Scheduler trait (default None; opt-in for schedulers with address recycling).
• Before cleanup graceful stops, filter the worker map against the scheduler's response; skip entries whose addresses are no longer live.
• Clear stale handles explicitly at every state transition where workers are known to be gone.
• Short-circuit wait_for_finish when the worker map is empty.

Follows: #1045 (receiver-side worker identity validation).