Unifying deep learning with item response theory: interval measurement, annotator debiasing, efficiency, and explainability