* removed `change_shape` flag; assume all models are converted from pytorch (or other) but are in BCHW order

Revert this change because rife models (even if not supported by chaiNNer) use a different order...

@adegerard Do you plan to finish this?

Sorry, I forgot.

I ended some tests: I have to check another time grayscale images.
I wonder whether I have to put back the possibility to change from BCHW to BHWC (onnx_auto_split function)
The reason is that is not implemented in pyTorch.
Except for rife onnx models, I found no other models with this property.

I have an implementation which supports this (not commited, and not tested because I currently have no model to test this)
What do you think?

I found no other models with this property.

Sorry that I didn't give context to this. This is common in models exported from tensorflow

Thanks for finishing it. I'll review it later today