Extend Argilla integration TextGeneration, Preference, and more

[alvarobartt]

Description

This PR adds the PreferenceToArgilla step that will basically push a dataset to Argilla for preference i.e. rating and rationale over a list of generated responses for a given instruction. The PreferenceToArgilla step works with both existing suggestions i.e. ratings and rationales and injects those as suggestions within each record, but is also helpful for datasets that want to be hand annotated by humans with no suggestions i.e. no ratings and rationales. Additionally, note that this task provides a seamless integration with UltraFeedback so that's easier than ever to upload the UltraFeedback generated and annotated data to Argilla.

Also some minor tweaks and improvements have been applied to PromptCompletionToArgilla renamed to TextGenerationToArgilla.

Besides that, this PR also adds the KeepColumn step which basically is useful for column sorting and filtering before the pipeline.run ends, and I found it very convenient when running diverse pipelines with unsorted outputs and producing more columns that the ones I wanted, so KeepColumns helps with that and also respects the order provided via the columns arg.

Additionally, also the docstrings have been fixed and aligned for both CombineColumns and KeepColumns, as well as adding the unit tests for both steps. Finally, this PR also runs codespell to fix some typos in the codebase.

Example

from distilabel.integrations.argilla import PreferenceToArgilla

if __name__ == "__main__":
    ...
    push_to_argilla = PreferenceToArgilla(
        name="push_to_argilla",
        api_url="<ARGILLA_URL>",
        api_key="<ARGILLA_API_KEY>",
        dataset_name="push_to_argilla",
        dataset_workspace="admin",
        num_generations=2,
    )
    ...connect(push_to_argilla)

    dataset = pipeline.run(...)

[gabrielmbmb]

LGTM!