Scalable Representation Learning in Linear Contextual Bandits with Constant Regret Guarantees