Benchmarking Robust Self-Supervised Learning Across Diverse Downstream Tasks