Mitigating Data Sparsity for Short Text Topic Modeling by Topic-Semantic Contrastive Learning