DP-Dueling: Learning from Preference Feedback without Compromising User Privacy