This MR aims to bring the support of the train/test mode flag, for the BatchNorm operator.
Two MRs will follow this one : one for the CPU backend and one for the CUDA backend.