From Clozing to Comprehending: Retrofitting Pre-trained Masked Language Model to Pre-trained Machine Reader