Group Robust Preference Optimization in Reward-free RLHF