Reparameterized LLM Training via Orthogonal Equivalence Transformation