[Exlv2] Allow setting "num_experts_per_token" in the WebUI

[NiriProject]

Description
For those who don't mind trading some speed and vram for some reduced perplexity, this setting is great for squeezing just that extra bit from mixtral.

Additional Context

The setting can be found in "q_mlp.cu" It used to be hardcoded to 2, but is now simple integer variable. I've been editing it manually, but it's probably possible to pass a user defined setting from the webui.

[oobabooga]

Added here #4955. Could you check if it works?

[github-actions]

This issue has been closed due to inactivity for 6 weeks. If you believe it is still relevant, please leave a comment below. You can tag a developer in your comment.

[Thireus]

@AncientZarko, do we have some metrics about perfs and PPL improvements versus num_experts_per_token used?

Edit: https://www.reddit.com/r/LocalLLaMA/comments/18la6ao/optimal_number_of_experts_per_token_in/ found this Reddit thread.