V-SEAM: Visual Semantic Editing and Attention Modulating for Causal Interpretability of Vision-Language Models