FlatQuant: Flatness Matters for LLM Quantization