Attention Condensation via Sparsity Induced Regularized Training