GRATH: Gradual Self-Truthifying for Large Language Models