Bypassing the Simulator: Near-Optimal Adversarial Linear Contextual Bandits