A2PO: Towards Effective Offline Reinforcement Learning from an Advantage-aware Perspective