A Mini-Block Natural Gradient Method for Deep Neural Networks