Nonconvex Factorization and Manifold Formulations are Almost Equivalent in Low-rank Matrix Optimization