UAlign: Leveraging Uncertainty Estimations for Factuality Alignment on Large Language Models