Refactor the language model interface

[rlouf]

The idea is to simplify calls to LLMs by encapsulating the prompt they are called with and providing a single entry points for all models in the library. This PR implements the following design:

import outlines.text as text


def ask_an_expert(question):

    @text.completion("openai/davinci", stops_at=["\n", "."])
    def find_expert(question):
        """
        Q: ${question}
        A: A good person to answer this question would be
        """

    @text.completion("openai/davinci")
    def get_answer(meta_prompt, expert):
        """
        ${meta_prompt}
        For instance, ${expert} would answer
        """

    expert, meta_prompt = find_expert(question)
    answer, _ = get_answer(meta_prompt, expert)

    return answer


answer = ask_an_expert("What is the Earth's diameter?")
answer.eval()

There are several advantages to this approach:

1. Encapsulates the prompt/LLM call logic so the rest of the code conveys the user's intent better;
2. Prompts don't get in the way of the code; This is aligned with the use of a templating language;
3. This make it easier to manage prompts and add them in version control: they're just functions.

Closes #43.