Improving cross-lingual language understanding with consistency regularization-based fine-tuning