Towards Better Model Understanding with Path-Sufficient Explanations