The In-Sample Softmax for Offline Reinforcement Learning