Robustness Gap in NLP Models for Vulnerability Descriptions: Benchmarking and Data Augmentation