Adversarial Self-Training Improves Robustness and Generalization for Gradual Domain Adaptation