RoLoRA: Fine-tuning Rotated Outlier-free LLMs for Effective Weight-Activation Quantization