Accurate and Data-Efficient Toxicity Prediction when Annotators Disagree