Improving Domain Generalization for Sound Classification with Sparse Frequency-Regularized Transformer