fix(gateway): clean up cached agents on shutdown (#11205)

[Tranquil-Flow]

What changed and why

On gateway shutdown, _stop_impl() called _finalize_shutdown_agents() which only cleaned up agents actively mid-turn (from _drain_active_agents()). Idle agents sitting in _agent_cache (LRU cache) were never cleaned up, so their MemoryProvider.on_session_end() hooks never fired — causing potential memory/resource leaks from providers that persist state on session end.

Fix: Add an explicit cache sweep in _stop_impl() after _finalize_shutdown_agents():

1. Acquire _agent_cache_lock
2. Snapshot all entries and clear the cache
3. Release lock
4. Call _cleanup_agent_resources() on each agent

The sweep is positioned before adapter disconnection so memory providers can still make network calls (e.g. persist memories) before adapters tear down. Lock is held only during snapshot+clear, not during slow cleanup iteration — matching the pattern in _sweep_idle_cached_agents.

Closes #11205

How to test

1. Start gateway with a memory provider configured (e.g. Mnemoria)
2. Chat with the agent so a session is cached
3. Stop the gateway
4. Verify on_session_end() fires for the cached agent (check memory provider logs)

Automated: 8 new tests in tests/gateway/test_shutdown_cache_cleanup.py:

• Cached agent's shutdown_memory_provider called on stop
• Cache cleared after shutdown
• Empty cache — no error
• Multiple cached agents all cleaned
• Exception from one agent doesn't prevent cleanup of others
• Non-tuple cache entries handled
• None cache entries skipped
• Double-presence (both running + cached) cleaned at least once

Platform tested

• macOS (Darwin 24.6.0)

[teknium1]

Merged via salvage PR #16207 (commit bf05b8f). Your commit was rebase-merged onto current main with authorship preserved in git log. Thanks for the fix!