Multi-Granularity Semantic Revision for Large Language Model Distillation