BERT Meets CTC: New Formulation of End-to-End Speech Recognition with Pre-trained Masked Language Model