RNN-based Online Learning: An Efficient First-Order Optimization Algorithm with a Convergence Guarantee