Only Pay for What Is Uncertain: Variance-Adaptive Thompson Sampling