MQuant: Unleashing the Inference Potential of Multimodal Large Language Models via Full Static Quantization