Leveraging Human Feedback to Scale Educational Datasets: Combining Crowdworkers and Comparative Judgement