A Minimalist Approach to Offline Reinforcement Learning