On the Complexity of Representation Learning in Contextual Linear Bandits