Inductive Bias of Gradient Descent for Exponentially Weight Normalized Smooth Homogeneous Neural Nets