A Simple and Optimal Approach for Universal Online Learning with Gradient Variations