Add Magpie and MagpieGenerator tasks

[gabrielmbmb]

Description

This PR adds two new tasks Magpie and MagpieGenerator based on Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing paper, which presents a new way to leverage chat template and the autoregressive capabilities of the instruct fine-tuned LLMs to generate instructions without seed data or a specific system prompt.

In addition, a new mixin MagpieChatTemplateMixin has been added that has to be used in combination with the aforementioned tasks. This mixin requires LLMs that are able to send the raw input (with special tokens, etc) to the LLM, so at this moment only the TransformersLLM, vLLM and InferenceEndpointsLLM uses the mixin and therefore are the only LLMs that can be used with the new tasks.

Finally, InferenceEndpointsLLM has been updated and the use_openai_client argument has been removed (and the use of the OpenAI client) as it was not used and didn't add any additional functionality that the huggingface_hub.AsyncInferenceClient doesn't offer. In addition, if tokenizer_id is not provided, then the chat_completion method from the huggingface_hub.AsyncInferenceClient is used instead of the text_generation. ⚠️ Even though with the new chat_generation methods a list of tools can be passed, we need to update the LLM.generate function signature and adapt the agenerate method so function calls can be returned too #782 (I'll tackle this in a different PR).

[github-actions]

Documentation for this PR has been built. You can view it at: https://distilabel.argilla.io/pr-778/

[codspeed-hq]

CodSpeed Performance Report

Merging #778 will not alter performance

_{Comparing magpie (cd3cc5d) with develop (86d4e80)}

Summary

✅ 1 untouched benchmarks

[plaguss]

LGTM! will review again once it's ready for review 👍