Add PairRM task for ranking responses

[plaguss]

Description

This PR adds PairRM to rank responses from different LLMs.

Example pipeline:

from typing import Any, Dict, Generator, List

from distilabel.pipeline.local import Pipeline
from distilabel.steps.base import RuntimeParameter, Step, StepInput
from distilabel.steps.generators.huggingface import LoadHubDataset
from distilabel.steps.combine import CombineColumns

from distilabel.steps.task.pair_rm import PairRM


class GenerateResponse(Step):
    @property
    def inputs(self) -> List[str]:
        return ["instruction"]

    @property
    def outputs(self) -> List[str]:
        return ["response"]

    def process(self, inputs: StepInput) -> Generator[List[Dict[str, Any]], None, None]:
        for input in inputs:
            input["response"] = "Response here."

        yield inputs


class RenameColumns(Step):
    rename_mappings: RuntimeParameter[Dict[str, str]]

    @property
    def inputs(self) -> List[str]:
        return []

    @property
    def outputs(self) -> List[str]:
        return list(self.rename_mappings.values())  # type: ignore

    def process(self, inputs: StepInput) -> Generator[List[Dict[str, Any]], None, None]:
        outputs = []
        for input in inputs:
            outputs.append(
                {self.rename_mappings.get(k, k): v for k, v in input.items()}  # type: ignore
            )
        yield outputs


if __name__ == "__main__":
    with Pipeline() as pipeline:
        load_hub_dataset = LoadHubDataset(name="load_dataset", batch_size=8)
        rename_columns = RenameColumns(
            name="rename_columns",
            input_batch_size=12,
        )
    
        text_generator_1 = GenerateResponse(
            name="generate_response_1", input_batch_size=16
        )
        text_generator_2 = GenerateResponse(
            name="generate_response_2", input_batch_size=16
        )
    
        load_hub_dataset.connect(rename_columns)
        rename_columns.connect(text_generator_1)
        rename_columns.connect(text_generator_2)

        # Will combine the outputs from the different LLMs (the candidates to rank)
        combine_columns = CombineColumns(
            name="combine_columns",
            merge_columns=["response"],
            output_merge_columns=["candidates"],
        )
        
        text_generator_1.connect(combine_columns)
        text_generator_2.connect(combine_columns)

        # Prepares the column names to the name expected from `PairRM`
        rename_for_pairrm = RenameColumns(
            name="rename_for_pairrm",
            input_batch_size=12,
        )
        combine_columns.connect(rename_for_pairrm)

        ranker = PairRM(name="pair_rm_ranker")
        rename_for_pairrm.connect(ranker)

        ds = pipeline.run(
            parameters={
                "load_dataset": {
                    "repo_id": "plaguss/test",
                    "split": "train",
                },
                "rename_columns": {
                    "rename_mappings": {"prompt": "instruction"}
                },
                "rename_for_pairrm": {
                    "rename_mappings": {"instruction": "input"}
                },
            }
        )
        print(ds)
        print(ds["pair_rm_ranker"][0])

Closes #433