Internal Representations of Vision Models Through the Lens of Frames on Data Manifolds