Scope Runs API by session key

[necoweb3]

This stores the creating request's X-Hermes-Session-Key for each /v1/runs run and enforces that owner key on the run follow-up endpoints.

Why

The API server allows multiple clients to share one API_SERVER_KEY while using X-Hermes-Session-Key as the per-user/channel boundary. POST /v1/runs already parses and forwards that key, but the follow-up endpoints only checked the bearer token and run_id. A caller with the same API key and a different session key could read another run's status, attach to events, resolve approvals, or stop the run if it knew the run_id.

Wrong-key requests now return 404 so the run's existence is not disclosed.

Changes

• Add a run_id -> X-Hermes-Session-Key owner map for API runs.
• Check the owner key in:
  • GET /v1/runs/{run_id}
  • GET /v1/runs/{run_id}/events
  • POST /v1/runs/{run_id}/approval
  • POST /v1/runs/{run_id}/stop
• Clean up the owner map with the existing run lifecycle cleanup.
• Add regression coverage for wrong-key status, events, approval, and stop attempts.

Tests

python -m pytest tests/gateway/test_api_server_runs.py -k "session_key or RunStatus or RunEvents or StopRun" -q --timeout-method=thread
19 passed, 7 deselected

[liuhao1024]

LGTM. The run-scoping mirrors the session-scoping pattern from #42460 cleanly:

• _authorize_run_session_key correctly returns None (allow) when the run has no owner key, preserving backward compatibility for unauthenticated callers.
• The owner key is stored alongside the run on creation and cleaned up on completion — no stale entries.
• Authorization is checked on all three run-specific endpoints (status, events, approval).
• The 404 response (not 403) is the right choice — it doesn't leak the existence of runs the caller can't access.

[necoweb3]

Rebased this PR onto current `origin/main` and reran the focused Runs API coverage.

Validation:

```text
python -m pytest tests/gateway/test_api_server_runs.py -q --timeout-method=thread
26 passed