Axiomatization of Gradient Smoothing in Neural Networks