fix(app-server): avoid blocking reconnect cleanup

[cbusillo]

Why

Rapid desktop quit/relaunch cycles could establish a replacement websocket while the app-server's single transport event loop was still synchronously draining RPCs from the disconnected client. In the reproduced failure, an old plugin/list remained active long enough for the replacement client's initialize handshake to exceed the desktop's 30-second deadline.

What changed

• Close the disconnected client's RPC gate immediately so queued work cannot start.
• Remove the dead connection from outgoing and thread-routing state before waiting for active RPCs.
• Run RPC draining and connection-scoped resource cleanup in tracked background tasks, while preserving graceful drain and forced-shutdown abort behavior.
• Add cleanup lifecycle logging and focused gate/cleanup task tests.
• Add a deterministic websocket regression that holds plugin/list, disconnects the first client, proves a replacement client initializes before the old RPC is released, and then verifies cleanup completion.

Validation

• just test -p codex-app-server connection_handling_websocket
• just test -p codex-app-server connection_cleanup
• just test -p codex-app-server connection_rpc_gate
• just test -p codex-app-server command_exec_process_ids_are_connection_scoped_and_disconnect_terminates_process
• just fix -p codex-app-server
• Signed macOS engine canary: five clean quit/relaunch cycles and five immediate relaunch cycles all initialized over websocket against the same supervised engine PID, with no stdio fallback.

The full codex-app-server suite completed 888/896 tests; the eight failures were in unrelated existing command exec, remote store, thread history/rollback, turn-start, and web-search tests.

Refs #464