On the Importance of Feature Separability in Predicting Out-Of-Distribution Error