Variance-Dependent Regret Lower Bounds for Contextual Bandits