Rethinking Optimal Transport in Offline Reinforcement Learning