Variation-aware Vision Token Dropping for Faster Large Vision-Language Models