RRM: Robust Reward Model Training Mitigates Reward Hacking