Provable Guarantees for Self-Supervised Deep Learning with Spectral Contrastive Loss