docs(quest): add realtime-pipeline Quest setup; fix np.trapz + cli glue

[AlanYWu]

Summary

• Add docs/running-on-quest.md adapted to the final-realtime-pipeline
  architecture (two WebSockets — ws://127.0.0.1:8765 predictions,
  ws://127.0.0.1:8766 markers), driven by python -m pipeline collect
  and python -m pipeline realtime, with FreeNavScene wiring and Path B
  adb reverse / LAN instructions for both ports. Replaces the old
  tools/bci_ws_stub_server.py flow on port 5000.
• Fix np.trapz (removed in NumPy 2.x) → np.trapezoid in
  eeg_to_meta/inference.py and ml_pipeline/features.py. Without this,
  pipeline realtime crashes on first inference under NumPy ≥ 2.0.
• pipeline/cli.py: drop the --mi-threshold / --blink-threshold
  flags and the override block — per-class thresholds stay hardcoded in
  the classifier model bundle's CLASS_THRESHOLDS. Also forward
  participant and sentinel mi_threshold=None, blink_threshold=None
  on the eeg_args Namespace so eeg_to_meta/main.py:run_pipeline
  doesn't AttributeError.

Verification

• python -m pipeline realtime --participant 5 --skip-personalize \\ --model ml_pipeline/models/realtime_models.pkl --mock boots the
  WebSocket on ws://127.0.0.1:8765 and emits stable predictions
  (Active class thresholds: {mi_*: 0.55, intentional_blink: 0.6}).

Test plan

• Unity FreeNavScene Editor Play → log shows
  [PredictionWebSocketClient] Connected to ws://127.0.0.1:8765
  and the server log flips to ws_clients=1.
• AutoMove scene + python -m pipeline collect --participant N --runs 1 --mock → markers reach the Python collector on :8766.
• Path B (APK on Quest): adb reverse tcp:8765 tcp:8765 and
  tcp:8766 tcp:8766, then re-verify both connects.