Deep Generative Pattern-Set Mixture Models for Nonignorable Missingness