Marrying Adapters and Mixup to Efficiently Enhance the Adversarial Robustness of Pre-Trained Language Models for Text Classification