Sample-Efficient Reinforcement Learning via Counterfactual-Based Data Augmentation