Learning from Invalid Data: On Constraint Satisfaction in Generative Models