From Uncertainty to Trust: Enhancing Reliability in Vision-Language Models with Uncertainty-Guided Selective Decoding