Universal Stagewise Learning for Non-Convex Problems with Convergence on Averaged Solutions