Enhancing Negation Awareness in Universal Text Embeddings: A Data-efficient and Computational-efficient Approach