Guided Online Distillation: Promoting Safe Reinforcement Learning by Offline Demonstration