Learning Constrained Optimization with Deep Augmented Lagrangian Methods