Towards Costless Model Selection in Contextual Bandits: A Bias-Variance Perspective