Prototypical Reward Network for Data-Efficient RLHF