Towards User-level Private Reinforcement Learning with Human Feedback