Add selective zero-copy DataFrame interchange

[cnpryer]

I thought I'd open this issue to surface the interest and serve as a place to coordinate if necessary.

The ask is to start opening up the protocol usage for any use-cases that are zero-copy compatible.

As it stands, if I understand, the state of converting between Pandas and Polars is:

• Convert to Pandas zero-copy with feat(python): Add option to use PyArrow backed-extension arrays when … #6756
• Convert from or to Pandas with copy from feat(python): Implement DataFrame Interchange Protocol through pyarrow #6581

Cannot utilize from_dataframe in a zero-copy manner. This will throw a NotImplementedError. We can refine this later by detecting which dtypes are present in the dataframe and determining if those can be handled in a zero-copy manner.
- #6581

I looked at this stuff a while ago because this feature specifically interests me. So I generally understand the level of effort for this ask. I'm in a spot where I'd like to migrate most of our Pandas usage to Polars. I've visited this topic a few times in the past year. The conversation is much easier with zero-copy options both ways. So I'd be very excited to see this feature implemented wherever it can be.

[stinodego]

I think @ghuls added this recently in #6844

We can probably still improve!

[cnpryer]

Of course @ghuls did lol. Such a machine. That's awesome.

I plan to migrate a lot of stuff to polars over the next month, so I'll give it a spin.

[stinodego]

I am re-opening this as I don't think the current from_dataframe implementation guarantees zero copy in the right way. See my comment here.

Let's see how this turns out - I think adding the proper safeguards to guarantee zero-copy is actually not too big a hurdle. I don't mind doing some work there.