Learning Robust Representations with Long-Term Information for Generalization in Visual Reinforcement Learning