Fairness in Reinforcement Learning with Bisimulation Metrics