SA-Attack: Improving Adversarial Transferability of Vision-Language Pre-training Models via Self-Augmentation.