BARReL: Bottleneck Attention for Adversarial Robustness in Vision-Based Reinforcement Learning