Embedded-model flows: Combining the inductive biases of model-free deep learning and explicit probabilistic modeling