[BUG] Coordinator leaks request/future state when submit_to_stage fails

[pikaxinge]

Summary

Coordinator keeps stale request state when entry-stage submission fails.

When submit_to_stage raises in _submit_request, the coordinator can retain:

• _requests[request_id]
• _completion_futures[request_id]

This leaves partially-created request state in memory.

Impact

• State accumulation in long-running services
• Increased memory/GC pressure under repeated submit failures
• Inconsistent request bookkeeping after failed submit

Repro (minimal)

1. Create a coordinator with one entry stage.
2. Make control_plane.submit_to_stage(...) raise an exception.
3. Call await coordinator.submit(request_id, request).
4. Observe leaked keys in _requests / _completion_futures.

Expected

On any submit failure, coordinator should fully rollback request bookkeeping.

Related PR

• Fix coordinator state leaks when submit fails #161