DiffSampling: Enhancing Diversity and Accuracy in Neural Text Generation