Implement FirstToKAhead Sampler

[kiranandcode]

Closes #407

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


@Template.define
def classify_genre(plot: str) -> MovieGenre:
    """Classify the movie genre based on this plot: {plot}
    Respond with exactly one of: action, comedy, drama, horror, sci-fi, romance"""
    raise NotImplementedError

plot = "A rogue cop must stop a terrorist group from detonating bombs across the city."
with handler(LLMLoggingHandler()), handler(OpenAIAPIProvider(openai.OpenAI())):
    with handler(KAheadSampler()):
        genre = classify_genre(plot)
    print(f"\nsampled genre: {genre}")

[jfeser]

The code as written seems fine, but I think we'd do better to eventually provide a more general batching interface. One option could be:

@defop
def sample(template: Template[P, T], n: int) -> Template[P, Sequence[T]]: ...

[jfeser]

This won't cancel ongoing work. Not sure if that's the desired semantics.

[kiranandcode]

yep, I figured raising exceptions might mess up the state, so best to just let the subprocesses run to completion.