Offline Multi-Agent Reinforcement Learning with Implicit Global-to-Local Value Regularization