Optimal Rates for Zero-Order Convex Optimization: The Power of Two Function Evaluations