Where the Model Looks: Uncovering Racial and Gender Bias in Vision-Language Models Through Visual Attribution