Towards Passive Safe Reinforcement Learning: A Comparative Study on Contact-rich Robotic Manipulation