Exploring the Manifold of Neural Networks Using Diffusion Geometry