Scalable imputation of genetic data with a discrete fragmentation coagulation process