Sparse Attention Post-Training for Mechanistic Interpretability