feat(tosa): sink tosa ops through tosa.concat

[gerion-amd]

Currently supports elementwise and reduce operations, additional to matmul and reshape (within a limited set).
Could be extended to other operations later.

[jorickert]

Thanks, especially for adding so many tests.

It looks generally good to me, but I will need to think about the reshape again, the current implementation seems to only handle some cases

[gerion-amd]

It looks generally good to me, but I will need to think about the reshape again, the current implementation seems to only handle some cases

This is correct. Only reshapes that extend or lessen the shape with dimensions of rank 1 will be transformed. This seemed to be enough for real-world tests, so I left it that way, with the possibility to handle more advanced cases later. Maybe I can add a comment explaining this.

[gerion-amd]

All comments should be addressed now.

[ehsan-toosi]

I did an high-level review and checked the transformation. Just one major comment which was using trait as you and Jonas already discussed as well.
I'll do the detailed code review if I find some time today.

[roberteg16]

Looks very nice! I have yet to finish the review of the reshape, I'll come back ASAP

[roberteg16]

LGTM, I would like just to have a few more reshape tests

[gerion-amd]

Changes from 0803705 to ba93cf8:

• Rename the variables with old and new prefix to beforeReshape and afterReshape to make it more specific that their position in the operation graph is meant
• Add documention of the reshape method
• Fix a bug for a reshape where the concat axis dimension is not of size 1. The approach is to also include the concat axis dimension size into the product and check for the same dimension before and afterwards.

[gerion-amd]

I added a second commit to have options to disable parts of the pass for better testing. This could be extended to optionally disable the relu operations mentioned above, if they get problematic.

[ehsan-toosi]

Thanks.

[roberteg16]

Thanks!