Input broadcasting for Add operator

For the ViT model, some "Add" operators require broadcasting of inputs. For input tensors A and B, these are the shapes cases for the ViT model:

A.shape == B.shape  # Only case currently supported 
A.shape == [1] or B.shape == [1] # Broadcasting is done on all dimensions
A.shape == [B.shape[-1]] or B.shape == [A.shape[-1]] # Broadcasting is done on all dimensions except the last one
A.shape == [B.shape[:-1], 1] or B.shape == [A.shape[:-1], 1] # Broadcasting is done on the last dimension

There are other Broadcasting cases supported by onnx and numpy but they are not needed for ViT.

added New Feature label

changed milestone to %aidge_core - v0.2.0

assigned to @hrouis

mentioned in merge request !65 (merged)

Solved in !65 (merged)

closed