What does BERT Learn from Arabic Machine Reading Comprehension Datasets?