Geometry-Aware Deep Congruence Networks for Manifold Learning in Cross-Subject Motor Imagery