Advancing Multimodal Large Language Models with Quantization-Aware Scale Learning for Efficient Adaptation