Position: Adopt Constraints Over Penalties in Deep Learning