fix(gateway): tear down cached agents on stop (salvage #12917)

[teknium1]

Gateway shutdown now shuts down memory providers on idle cached agents, not just mid-turn ones. Closes leak of Hindsight/Honcho/Mnemoria clients + terminal sandboxes + browser daemons at systemd stop.

Salvage of #12917 by @Tranquil-Flow. Single contributor commit cherry-picked onto current main; author attribution preserved via rebase merge.

Root cause

_drain_active_agents snapshots only _running_agents. Agents that had ended their turn and were living in _agent_cache (warm cache / idle TTL sweep pool) never received shutdown_memory_provider() or close(), so their memory-provider network clients, terminal sandboxes, and browser daemons stayed alive until process exit.

Changes

• gateway/run.py _stop_impl: after _finalize_shutdown_agents(active_agents), sweep _agent_cache under _agent_cache_lock, then call _cleanup_agent_resources on each entry (handles tuple and bare-agent forms).
• tests/gateway/test_shutdown_cache_cleanup.py: 8 new tests covering memory-provider shutdown ordering, cache-cleared-after, exception resilience, multiple agents, plain/tuple/None entries, and the running+cached overlap case.

Validation

	Before	After
Shutdown path iterates	_running_agents only	_running_agents + _agent_cache
Targeted tests	—	8/8 pass (1.1s)

Closes #12917 (original PR, merged via salvage). Addresses #11205.