fix(gateway): bound SQLite file descriptor usage

[injooinjoo]

Summary

• reuse the gateway SessionStore's existing SessionDB instead of creating a second long-lived instance
• cap per-thread WAL reader connections at 16 and fall back to the locked writer connection after the cap
• make worker-created read connections cross-thread closable so SessionDB.close() actually releases their descriptors
• de-duplicate SessionDB shutdown when the runner and store share one instance

Reproduction

A long-lived macOS gateway with launchd's soft RLIMIT_NOFILE=256 reached 252 numeric descriptors. 145 were state.db / WAL descriptors. GatewayRunner created two SessionDB instances, and each retained one WAL reader connection per worker thread. In addition, readers were opened with SQLite's default check_same_thread=True, so the owner-thread close loop raised ProgrammingError and silently left worker-created descriptors open.

Verification

• scripts/run_tests.sh tests/test_hermes_state.py tests/test_session_db_read_path_split.py tests/gateway/test_session.py tests/gateway/test_gateway_shutdown.py tests/gateway/test_runner_session_db_fd_budget.py -q
• 257 tests passed on current main
• 100 short-lived reader-thread stress: 16 tracked readers at peak; process FD count returned from 39 to 4 after SessionDB.close()
• local launchd gateway after hotfix restart: 39 numeric FDs / 6 state.db FDs, stable for 60 seconds; Slack auth remained healthy