Can Pretext-Based Self-Supervised Learning Be Boosted by Downstream Data? A Theoretical Analysis