P4Q: Learning to Prompt for Quantization in Visual-language Models