Bugs in Machine Learning-based Systems: A Faultload Benchmark