Bayesian Pool-based Active Learning With Abstention Feedbacks