Building Chinese Biomedical Language Models via Multi-Level Text Discrimination