50 Years of Test (Un)fairness: Lessons for Machine Learning