Look Before You Leap: Safe Model-Based Reinforcement Learning with Human Intervention