Using Deep Learning to Identify Initial Error Sensitivity for Interpretable ENSO Forecasts