Cortex_m backend: Support channels-broadcasting for ADD/MUL

[AdrianLundell]

Adds support for broadcasting in the special case where one input contains only channels which are broadcasted onto every spatial element of the other tensor, e.g. [1,C,1,1] + [N, C, H, W] for channel-last tensors.

This is a needed for mobilenet_v3.

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/16131

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 2 New Failures

As of commit 984465d with merge base 488d761 ():

NEW FAILURES - The following jobs have failed:

• pull / android / run-emulator (gh)
  Timeout waiting for emulator to boot.
• trunk / test-arm-cortex-m-size-test (zephyr-preset) / linux-job (gh)
  RuntimeError: Command docker exec -t 6572ce7c7b755ec98601d5f15ef92d8cc752aa2eae61f31b1e943faf4e39be4c /exec failed with exit code 1

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[psiddh]

This PR does not implement general broadcasting semantics, but rather adds support for a very specific broadcasting case required for MobileNet v3 (MV3).
In general, should we aim to support full broadcasting semantics?

[AdrianLundell]

My two cents is that full broadcasting support takes engineering effort, code size, and likely performance tradeoffs, and on the other hand the selection of networks which makes sense to run on MCUs is fairly limited and might not use a lot of broadcasting.

So in my view only implementing what is used makes sense here, until we see a large need for full broadcasting support (if ever).

[psiddh]

My two cents is that full broadcasting support takes engineering effort, code size, and likely performance tradeoffs, and on the other hand the selection of networks which makes sense to run on MCUs is fairly limited and might not use a lot of broadcasting.

So in my view only implementing what is used makes sense here, until we see a large need for full broadcasting support (if ever).

Seems reasonable, if more usecases / models start requiring general broadcasting, we could look at supporting it more broadly