Optimistic Regret Bounds for Online Learning in Adversarial Markov Decision Processes