Add list manipulation expressions

[joseph-isaacs]

What expressions do we need to operator for DType::List.

I am currently thinking that we want map_list(func, array). The question becomes how to be handle function in expressions.

If we had a map_list(get_item("a", _), array) where the array has dtypelist({a: i32, b: u8}) expression this would extract the column "a" for each list element returning an array of dtype list(i32)

[caicancai]

I'm curious what conventions this list's type inference follows. I'm not sure if there will be a type inference problem 🤔

[godnight10061]

For the motivating case (map_list(get_item("a", _), list<struct{a,b}>)), I wonder if we can avoid higher‑order funcs entirely by defining get_item("a", list<struct{a: T, ...}>) -> list (and same for FixedSizeList).
Type inference is then purely structural: preserve the list container (validity/offsets), keep null elements null, and propagate the field’s own validity. Error at planning time if the element dtype isn’t a struct or the field is missing.
I have PR #6012 that does this + adds a unit regression and a small write→scan smoke test for items.a.

[joseph-isaacs]

I really think what we want to follow all other query engines here and have map(lambda x: get_item("a", x), array).

The question becomes do we have monomorphised functions map_list or a generic one map. (I prefer map)

[godnight10061]

Agreed. We do not have lambda/HOF map in exprs yet, so this is a pragmatic list.field fix now; it can become sugar for map later. Should we prioritize adding generic map next?

[joseph-isaacs]

Sorry for the delay.

How about we add a temporary unstable expression get_item_list(Col, array) that does what you wish?

We will may remove this expression at anytime, likely once we have the higher order function, map which will replace it.

[joseph-isaacs]

We can add a temporary UNSTABLE get_item_list expression

This should have the semantics

get_item_list(Col, array) == map(lambda x: get_item(Col, x), array)

[godnight10061]

I replaced vortex.map with a temporary vortex.get_item_list(col, array) (as sugar for a future map) and updated list pushdown accordingly; the existing write->scan regression for items.a and the unit tests pass, with no new deps.