Fix request settings and stop words in HF provider

[JonasHelming]

fixed #14503

What it does

• Huggingface provider respects request settings
• Added Huggingface default settings (the stop words are very common)
• Remove all configured stop words from response

How to test

• Set up a HuggingFace Model
• e.g.: "Qwen/Qwen2.5-Coder-32B-Instruct"
• Configure it for the universal chat agent
• Ask it a question
• Observe that always end with '<|im_end|>'
• Change HF defaults in the model:

   protected getDefaultSettings(): Record<string, any> {
        return {
            max_new_tokens: 2024,
            stop: ['<|endoftext|>', '<eos>', '<|im_end|>']
        };
    }

• Observe that it does not end with '<|im_end|>' anymore
• Turn streaming of in the HuggingFaceModel
• Same test
• Turn streaming on in the HuggingFaceModel
• Remove default override
• Add settings to the universal-chat-agent

   protected override getLlmSettings(): { [key: string]: unknown } | undefined {
      return {
         max_new_tokens: 2024,
         stop: ['<|im_end|>'],
      };
   }
   

• Observe that it does not end with '<|im_end|>' anymore
• and without streaming

Follow-ups

• We must think about how we want to allow the user configure these settings (per model, per provider, per agent, per prompt?)
• In particular for the stop words: As HuggingFace does not replace them, I think we should as a first step allow to configure stop words for HuggingFace Models only. This way HF gets even usable.
• If we start to really use the settings, we need a compatibility check for all provider APIs. Most parameters are valid for all of them, but not all of them

Review checklist

• As an author, I have thoroughly tested my changes and carefully followed the review guidelines

Reminder for reviewers

• As a reviewer, I agree to behave in accordance with the review guidelines

[sdirix]

Works for me ❤️