Chaining Meets Chain Rule: Multilevel Entropic Regularization and Training of Neural Nets