Leveraging Reward Consistency for Interpretable Feature Discovery in Reinforcement Learning