Add support for classifying objects and returning indices

[jlowin]

result = marvin.classify('book', ['a thing you read', 'a thing you sing'], return_index=True)
assert result == 0

Above is a simple example but this is useful when you want to classify objects that aren't already strings. Getting the index makes it easy to match the object to the selected string description

In testing the above, I discovered and fixed (arguably) a bug that cleans up this DX even more. You can now pass objects as labels and get the object back (assuming it can automatically be turned into a string by Pydantic). Previously they would be converted into (bad) string representations, which were returned as the result of the classifier.

class Person(BaseModel):
    name: str
    age: int

p1 = Person(name="Alice", age=30)
p2 = Person(name="Bob", age=25)
p3 = Person(name="Charlie", age=35)

result = marvin.classify("a person in wonderland", [p1, p2, p3])
assert result is p1

[zzstoatzz]

interesting!