Distributionally Robust Model-Based Offline Reinforcement Learning with Near-Optimal Sample Complexity