Tight First- and Second-Order Regret Bounds for Adversarial Linear Bandits