feat: moe kernel tuning

[llcnt]

Description

This PR is inspired from vLLM benchmarks (the benchmark_config fn is copied from here) and enable one to tune the MoE (triton) kernel used in vllm.
This new algorithm MoeKernelTuner does not modify the model. It generates a tuned configuration that is saved in:

• the vllm configs folder (so that using the model on the same gpu afterward makes vllm use this optimized config);
• the RedhatAI kernel folder in the hf hub (so that using the moe kernels from the kernels lib will make use of the optimized config);
• the smash_config (to be saved and later re-used without waiting for tuning).

The core modification are in:

• the new moe_kernel_tuner.py file (...);
• the smash_config.py file (adding new artifacts for saving any additional dict into the smashconfig);
• the load.py file (for re-svaing the tuned config inside vllm/hf cache when loading a smashed model).

TODO:

• clean PR description
• add model pruna-test/qwen3_coder_tiny for unit tests

Related Issue

Fixes #(issue number)

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

How Has This Been Tested?

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

Additional Notes

Notebook for testing with vllm is available here

[github-actions]

This PR has been inactive for 10 days and is now marked as stale.