Data-Free Quantization via Mixed-Precision Compensation without Fine-Tuning