Understanding how Differentially Private Generative Models Spend their Privacy Budget