Revisiting Design Choices in Offline Model Based Reinforcement Learning