Add backward() for implemented operators on CUDA

Add the backward() implementation for currently supported operators on backend_cuda:

• Add
• AvgPool
• BatchNorm
• FC
• GlobalAvgPool
• MaxPool
• Pad
• Sub
• Reshape

assigned to @hrouis

Tentative implementation of backward functions for ReLU, Sigmoid and Tanh in this merge request

added StatusWork in Progress label

mentioned in merge request !25 (merged)

changed the description

changed the description

added StatusReview Ready label and removed StatusWork in Progress label

closed