Feedback Alignment Meets Low-Rank Manifolds: A Structured Recipe for Local Learning