Bandit Smooth Convex Optimization: Improving the Bias-Variance Tradeoff