Implement training mode for batch normalization

[Swopper050]

Postponed in #186 because the operator uses optional outputs, which makes the implementation quite confusing and batch normalization in train mode is not used for inference, which is the focus of this package.