Attention, Please! PixelSHAP Reveals What Vision-Language Models Actually Focus On