Model-based Offline Reinforcement Learning with Local Misspecification