[ENH] improved polars support - native logic

[fkiraly]

Issue about next steps in extending polars support in skpro.

Background: skpro already supports both lazy and eager polars containers, as polars_eager_table and polars_lazy_table.

However, key limitations:

• support is currently exclusively via conversion back/forth
• the round trip to numpy or pandas breaks the lazy chain

So, suggested next steps, using only the eager type:

1. try polars input/output with a few estimators, fit / predict only. Add tests in a dedicated polars test file, non-systematic for the start.
2. sklearn now supports polars - so, we should try to pass on polars frames in some of the estimators, extending the X_inner_mtype to both pandas and polars (in a list of str, with polars_eager_table). Example: GaussianProcess
3. if that works well, we should extend other regresssors wrapping sklearn in the same way.
4. next I would work on Pipeline, a composite. The Pipeline is native to skpro.

Once the above seems to work, I would look at the lazy` type.
Here, we should use the boilerplate layer to make lazy state changes.
Design will follow when we are there, ideas appreciated.

FYI @julian-fong.

[julian-fong]

Linking 6381 as discussion and coordination thread