QA-LoRA: Quantization-Aware Low-Rank Adaptation of Large Language Models