A Quantitative Geometric Approach to Neural-Network Smoothness