bring back map for AIFunction

[zzstoatzz]

and also ensures the thing we get back from decorating a function is an AIFunction not a partial of a wrapper of the function

this map implementation should be compatible with the existing tests

Example

In [1]: from marvin import ai_fn

In [2]: from typing_extensions import Literal

In [3]: Tag = Literal['bug', 'docs', 'enhancement']

In [4]: @ai_fn
   ...: def labels(content: str) -> set[Tag]:
   ...:     '''select an appropriate set of tags for the content'''
   ...:

In [5]: labels
Out[5]: AIFunction(fn=<function labels at 0x1096b0860>, environment=None, prompt='Your job is to generate likely outputs for a Python function with the\nfollowing signature and docstring:\n\n{{_source_code}}\n\nThe user will provide function inputs (if any) and you must respond with\nthe most likely result, which must be valid, double-quoted JSON.\n\nuser: The function was called with the following inputs:\n{%for (arg, value) in _arguments.items()%}\n- {{ arg }}: {{ value }}\n{% endfor %}\n\nWhat is its output?', name='FormatResponse', description='Formats the response.', field_name='data', field_description='The data to format.', render_kwargs={}, create=None)

In [6]: labels.map(["eek a bug in the docs", "i have an idea"])
Out[6]: [{'bug', 'docs'}, {'enhancement'}]