Model-Based Offline Reinforcement Learning with Pessimism-Modulated Dynamics Belief