Selective "Selective Prediction": Reducing Unnecessary Abstention in Vision-Language Reasoning