Adds Modal run and check and server

[tcapelle]

Thanks for the great repo.

I added a simple run_and_check modal so we can run this on any hardware, also added a server in modal so we can serve the run_and_check and use it on a GRPO pipeline.

• Use uv as dependency manager and move to a more modern prproject.toml
• Rework imports and file locations, moved to src/kernelbench to make editable install robust and best practice.
• Use absolute path's everywhere, removed unused imports
• I am also working on having observability on the generation part, added some weave traces here and there while I figure how this part works. I feel we could simplify a lot by using something like litellm and not handling LLM providers one by one.

[willhu-jpg]

Hi Thomas, thank you for the contribution! We're taking a look at your pull request.

[tcapelle]

I can split the PR, let's book some time and chat about how can I help you guys. For the time being I moved to work on TritonBench instead.