Modify metaclass to allow DataFrame[Foo] type propagation

[thomasaarholt]

The main goal of this was to enable the following pyright type annotations:

import patito as pt
import polars as pl

class Foo(pt.Model):
    name: str


df1 = Foo.DataFrame({"name":["Bean"]})
df2 = pt.DataFrame({"name":["Bean"]}).set_model(Foo) # Type of "df" is "DataFrame[Unknown]"

Before

df1 and df2 report: # Type of "df" is "DataFrame[Unknown]"

After

df1 and df2 report: # Type of "df" is "DataFrame[Foo]"

Discussion

This allows us to robustly pass around dataframes that have the patito Model embedded in the type annotation.
I am still a bit unsure on how I feel about a df being a DataFrame[Foo] before .validate() but after .set_model(). @dsgibbons and @JakobGM, any thoughts on this?

[dsgibbons]

If Foo.DataFrame and set_model are supposed to be analogous to creating a Pydantic model, shouldn't we be running validate before returning the DataFrame? This would make patito function more like Pydantic, and, as a bonus, we then know that the resulting table is actually of type Foo. I guess the main issue is that validate may be quite expensive, but wouldn't everyone using this library want to run validate anyway?

[gab23r]

I don't think the validation should be done in set_model under the hood, I don't think that it a problem either to have the DataFrame type without the validation. Maybe we can add the validate=False argument to set_model to make it clearer.

In any case, I think having the typing correct if a HUGE quality of life improvement, is there other things bloking this PR ?

[thomasaarholt]

I think I'll merge this one as-is, and we can have a think about further improvements going ahead.