Harmonic Loss Trains Interpretable AI Models