Taming False Positives in Out-of-Distribution Detection with Human Feedback