CROP: Conservative Reward for Model-based Offline Policy Optimization