MirrorCheck: Efficient Adversarial Defense for Vision-Language Models