Enhancing Spectral Graph Neural Networks with LLM-Predicted Homophily