Learning Thresholds with Latent Values and Censored Feedback