Between Stochastic and Adversarial Online Convex Optimization: Improved Regret Bounds via Smoothness