Improving Natural Language Understanding for LLMs via Large-Scale Instruction Synthesis