The Pareto Frontier of model selection for general Contextual Bandits