Model-based Offline Reinforcement Learning with Count-based Conservatism