Learning Fair Division from Bandit Feedback