RAMAC: Multimodal Risk-Aware Offline Reinforcement Learning and the Role of Behavior Regularization