Implemented ReducedKAheadSampler

[kiranandcode]

This PR builds on top of #412. Often we would like to use an LLM to generate types that can not be hashed, or whose outputs do not lie within a small enough set such that voting is likely to complete in any reasonable amount of time (think generating anything with a string component, or callables). Yet these outputs may still have some "essense" in a smaller set which would still provide useful signal for voting.

This PR implements a ReducedKAheadSampler which allows the user to provide a reducer function which maps LLM generated outputs to a reduced key-set for voting, and then runs the usual k-ahead sampling algorithm.

class MovieGenre(str, Enum):
    ACTION = "action"
    COMEDY = "comedy"
    DRAMA = "drama"
    HORROR = "horror"
    SCIFI = "sci-fi"
    ROMANCE = "romance"

@dataclass(frozen=True)
class MovieClassification:
    genre: MovieGenre
    explanation: str = Field(..., description="explanation for the given movie classification")

@Template.define
def classify_genre(plot: str) -> MovieClassification:
    """Classify the movie genre based on this plot: {plot}"""
    raise NotImplementedError


plot = "A rogue cop must stop a terrorist group from detonating bombs across the city."

with handler(LLMLoggingHandler()), handler(OpenAIAPIProvider(openai.OpenAI())):

    def reducer(classification: MovieClassification) -> MovieGenre:
        return classification.genre
    def select_best(results: list[MovieClassification]) -> MovieClassification:
        return max(results, key=lambda res: len(res.explanation))
    
    sampler = ReducedKAheadSampler(reducer, select_best=select_best)
    classification = handler(sampler)(classify_genre)(plot)
    print(classification)

[kiranandcode]

Orthogonal design consideration: I don't feel like KAheadSampler as an interpretation for Template.__call__ really makes sense. In particular, because it overrides the semantics for Template.__call__ for all Template.calls when the algorithm only works for one particular Template.__call__.

[jfeser]

I think the fact that we need separate Template.__call__ handlers for this and the previous voting algorithm and that handling Template.__call__ is too coarse suggests that we should build simpler batching functionality and implement voting on top.