Riemannian Manifold Learning for Stackelberg Games with Neural Flow Representations