[FEA] Support rolling-by expressions inside ``over()`` in cuDF-Polars

Grouped rolling-by expressions inside `over(...)` are not currently supported by cudf-polars. We should add both in-memory and streaming support for this expression family.

This issue should cover the common `rolling_*_by(...).over(...)` variants:

- [ ] `rolling_sum_by(...).over(...)`
- [ ] `rolling_mean_by(...).over(...)`
- [ ] `rolling_min_by(...).over(...)`
- [ ] `rolling_max_by(...).over(...)`

**Example**
```python
def test_rolling_sum_by_over(engine: GPUEngine) -> None:
"""``rolling_sum_by(...).over(...)``.

Status
------
- In-memory NOT supported
- Streaming NOT supported
"""
lf = pl.LazyFrame(
{
"g": ["A", "A", "A", "B", "B"],
"ts": [
"2025-01-01 09:00:00",
"2025-01-01 09:01:00",
"2025-01-01 09:04:00",
"2025-01-01 09:00:00",
"2025-01-01 09:03:00",
],
"x": [10.0, 20.0, 30.0, 40.0, 50.0],
}
).with_columns(pl.col("ts").str.to_datetime())

q = lf.select(
pl.col("x").rolling_sum_by("ts", window_size="5m").over("g").alias("rs")
)
assert_gpu_result_equal(q, engine=engine)
```

Start with the example test_rolling_sum_by_over and inspect the existing cuDF-Polars implementations and tests for rolling_*_by expressions and over(...). Verify both in-memory and streaming execution for rolling_sum_by, rolling_mean_by, rolling_min_by, and rolling_max_by, using assert_gpu_result_equal as the completion check.

Written by the indexing model from the issue text.