Bayesian Active Learning With Abstention Feedbacks