Why Physics Still Matters: Improving Machine Learning Prediction of Material Properties with Phonon-Informed Datasets