FoPru: Focal Pruning for Efficient Large Vision-Language Models