[FEA] Support outer null checks for Array columns in cudf-polars

[0guban0v]

Description

cudf-polars currently supports passing fixed-size Array columns through GPU. This feature request extends that support to is_null() and is_not_null() on Array columns.

Reproducer

import polars as pl

df = pl.DataFrame(
    {
        "embedding": pl.Series(
            [
                [1.0, 2.0],
                None,
                [None, None],
                [3.0, None],
            ],
            dtype=pl.Array(pl.Float32, 2),
        )
    }
)

query = df.lazy().select(
    pl.col("embedding").is_null().alias("is_null"),
    pl.col("embedding").is_not_null().alias("is_not_null"),
)

print(
    query.collect(
        engine=pl.GPUEngine(
            executor="streaming",
            raise_on_fail=True,
        )
    )
)

Output

NotImplementedError: Query execution with GPU not possible: unsupported operations.

Expected

Query should run on GPU and match Polars:

shape: (4, 2)
┌─────────┬─────────────┐
│ is_null ┆ is_not_null │
│ ---     ┆ ---         │
│ bool    ┆ bool        │
╞═════════╪═════════════╡
│ false   ┆ true        │
│ true    ┆ false       │
│ false   ┆ true        │
│ false   ┆ true        │
└─────────┴─────────────┘

Only outer validity of each Array row should be checked. [None, None] and [3.0, None] are valid Array rows.

Scope after Polars optimization

Polars supports all of expressions below. cudf-polars receives optimized plan, which may contain different operation from original Python expression. This issue covers direct outer null checks and expressions that preserve those checks after optimization.

Python expression	Optimized expression in Polars 1.44	This issue
a.is_null().cast(pl.Int8)	a.is_null().strict_cast(pl.Int8)	Supported
df.drop_nulls("a")	Filter on a.is_not_null()	Supported
a.is_null().sum()	a.null_count()	Out of scope
a.is_not_null().sum()	a.len() - a.null_count()	Out of scope
a.is_null().any()	a.has_nulls()	Out of scope
a.is_not_null().all()	a.has_nulls().not()	Out of scope