feat: reduce nb experts per token in moe architectures

[llcnt]

Description

This PR is a little tool that only acts on MoE models (for LLMs and Hunyuan3Image for now) by reducing the number of experts that are trigered for each token.
All models have been trained on a fix amount of active experts per token, and decreasing this number alter the output of the model. This idea was tested on Hunyuan3Image, gptoss_120b (for hunyaun3 image, (default is 8 out of 128 experts) 1 and 2 give very weird images: 4 experts seems ok, and yields 15% speedup. For gptoss120b, (default is 4 experts) 1 and 2 give very weird texts, and yields no speedup), but is applicable to any MoE, eg Mixtral, QwenNext, etc.

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 to test the new feature available here.

[cursor]

Comment @cursor review or bugbot run to trigger another review on this PR

[llcnt]

@cursor review

[simlang]

Love it, super straightforward algorithm. commented on some higher level stuff.