Extending Stan for Deep Probabilistic Programming