Bias no more: high-probability data-dependent regret bounds for adversarial bandits and MDPs