Use Jinja filters instead of a tool decorator

[rlouf]

We can now use custom Jinja filters to get function names, descriptions, sources and signatures inside prompt templates:

import outlines.text as text


@text.prompt
def a_prompt(question: str, fn: Callable):
    """{{question}}

    COMMANDS:
    {{fn|name}}
    {{fn|description}}
    {{fn|signature}}
    {{fn|source}}
    """

This removes the need for the @outlines.tool decorator defined in #64. We also add a filter to get the simplified schema of a Pydantic model:

import outlines.text as text
from pydantic import BaseModel, Field

class Joke(BaseModel):
    joke: str = Field(description="a good joke")
    explanation: str

@text.prompt
def a_prompt(question: str, response_format: type(BaseModel)):
    """{{question}}

    OUTPUT FORMAT:
    {{response_format|schema}}
    """

a_prompt("Can you tell me a good joke?", Joke)
# Can you tell me a good joke?
#
# OUTPUT FORMAT:
# {
#    "joke": "a good joke", 
#    "explanation": "<explanation>"
# }

Closes #68