Q-Supervised Contrastive Representation: A State Decoupling Framework for Safe Offline Reinforcement Learning | Read Paper on Bytez