Long-term Safe Reinforcement Learning with Binary Feedback