[FEA] Support `filter` in groupby/rolling for cudf-polars

```python
import numpy as np
import polars as pl

lf = pl.LazyFrame(
{
"a": [None, 2., 3.],
"b": [1, 1, 2],
}
)
with pl.Config() as cfg:
cfg.set_verbose(True)
print(lf.group_by('b').agg(pl.col('a').filter(pl.col('a')>2).sum()).collect(engine='gpu'))

"""
The errors were:
- NotImplementedError: No support for in groupby/rolling
return wrap_df(ldf.collect(engine, callback))
keys/aggregates are not partitionable: running default HASH AGGREGATION
shape: (2, 2)
┌─────┬─────┐
│ b ┆ a │
│ --- ┆ --- │
│ i64 ┆ f64 │
╞═════╪═════╡
│ 2 ┆ 3.0 │
│ 1 ┆ 0.0 │
└─────┴─────┘
"""
```

The cuDF version is '25.06.00a348'.