Ask One More Time: Self-Agreement Improves Reasoning of Language Models in (Almost) All Scenarios