Revisiting Bisimulation Metric for Robust Representations in Reinforcement Learning