Explaining medical AI performance disparities across sites with confounder Shapley value analysis