Low-rank Orthogonalization for Large-scale Matrix Optimization with Applications to Foundation Model Training - Databubble