mnist-inference-web wgpu matmul returns all zeros

[laggui]

Describe the bug
The wgpu matmul implementation returns all zeros when running in the browser.

Also reported by a discord user: https://discord.com/channels/1038839012602941528/1059209073784008835/1273995743790628876

To Reproduce
Run the mnist-inference-web example built for wgpu.

The scores returned are always the same for all digits.

Upon further debugging, the fc1 layer always returns zeros because of the matmul issue (tiling2d implementation w/ cube). Note that using the simple matrix multiply works (and cmma is not available in this case).

Expected behavior
The results from the matmul should not be zero, and thus the scores correct.

[nathanielsimard]

This should fix it: tracel-ai/cubecl#81

[laggui]

Just tested, this was fixed when updating to the latest cubecl version to include the fix in #2181