Fast Global Convergence for Low-rank Matrix Recovery via Riemannian Gradient Descent with Random Initialization