MG-LLaVA: Towards Multi-Granularity Visual Instruction Tuning