From Self-Supervised Speech Models to Mixture-of-Experts for Robust Anti-Spoofing