Disturbance Injection under Partial Automation: Robust Imitation Learning for Long-horizon Tasks