Universal and data-adaptive algorithms for model selection in linear contextual bandits