Avoiding the Price of Adaptivity: Inference in Linear Contextual Bandits via Stability