Beyond $\mathcal{O}(\sqrt{T})$ Regret for Constrained Online Optimization: Gradual Variations and Mirror Prox