Rethinking Pruning for Vision-Language Models: Strategies for Effective Sparsity and Performance Restoration