Predicting discrete-time bifurcations with deep learning