Improving Zero-shot Multilingual Neural Machine Translation by Leveraging Cross-lingual Consistency Regularization