Size and Depth Separation in Approximating Benign Functions with Neural Networks