From Attention to Atoms: Spectral Dictionary Learning for Fast, Interpretable Language Models