Predicting high-magnitude, low-frequency crop losses using machine learning: An application to cereal crops in Ethiopia