fix: capture a deployment logs before termination removes them

[Priyanshu-u07]

TerminalLogRepository.save() had no callers, so GET /logs/{deployment_id}/persisted returned nothing for every deployment on every provider.

The capture now runs in terminate_deployment_core, before unload_model stops the engine — afterwards the pod is gone and the logs with it. It is best-effort and bounded at 10 seconds, so a slow or unreachable provider cannot stall or fail a termination. The tail is capped at 500 lines because the rows are written inside the termination transaction.

FAILED and STOPPED are deliberately not wired, and the docstring now says so. STOPPED is set in a finally block after cleanup has already recycled the nodes, so it needs its own pre-teardown capture point inside worker.py. Most FAILED transitions are provisioning failures where no engine ever started.

Also removes the dangling TerminalLogRepository construction in worker_main.py that the issue names.

Closes #360