Warm-starting Contextual Bandits: Robustly Combining Supervised and Bandit Feedback