Decomposed evaluations of geographic disparities in text-to-image models