Draft: Use setBackend instead of ConvertToCpu workaround

Change the convertion to backend 'cpu' from a custom function converToCpu in the database to the generic setBackend().

Content :

• Remove the call to convertToCpu in database, its definition & tests

However it raises a problem on the layout of the data between different backends. In the current code the resulting tensor with backend 'cpu' have layout [N,H,W,C] instead of [N,C,H,W].