Reproducibility in Multiple Instance Learning: A Case For Algorithmic Unit Tests