Enhancing Efficiency of Safe Reinforcement Learning via Sample Manipulation