feat: add nw.Struct , nw.List, and nw.Array dtypes

[MarcoGorelli]

Polars has these natively

In pandas, they're available after pandas 2.0.0 with PyArrow dtypes, e.g.:

In [21]: pd.Series([[1, 2], [3, 4, 5]], dtype=pd.ArrowDtype(pa.list_(pa.int64())))
Out[21]:
0      [1 2]
1    [3 4 5]
dtype: list<item: int64>[pyarrow]

For pandas, I think it's OK to require pyarrow when casting to nw.List / nw.Struct / nw.Array

[FBruzzesi]

For motivation... plotly has the following aggregation in group by:

def aggfunc_discrete(x):
    uniques = nw.col(x).unique()
    if uniques.len() == 1:
        return uniques[0]
     else:
        return "(?)"

In the current status it would be hard for us to replicate. However if we manage to add List type and namespace then we could approach it as:

(df
.group_by("a")
.agg(nw.col("b").unique())
.select(
    nw.when(nw.col("b").list.len() == 1).then(nw.col("b").list.first()).otherwise("(?)")
 )
)

It's late so maybe I am missing something simpler than this snippet.

I think we should figure out how to do it in pandas without pyarrow backend via duck typing and raise a type/attribute error if we are not able to perform a list computation.

[MarcoGorelli]

i'm not sure about allowing non-aggregations in Narwhals, but regarding the unique issue raised here, I have a suggestion in #1076 (comment)

[MarcoGorelli]

These are all available in Narwhals now so we can close this