Noisy Gradient Descent Converges to Flat Minima for Nonconvex Matrix Factorization