Offline RL Policies Should be Trained to be Adaptive