Understanding and Addressing the Pitfalls of Bisimulation-based Representations in Offline Reinforcement Learning