Video Prediction Models as Rewards for Reinforcement Learning