Counterfactually Fair Reinforcement Learning via Sequential Data Preprocessing