We need evals for interpretability tools, not just prettier plots