Towards Safe Reinforcement Learning with a Safety Editor Policy