[BUG] Comparisons involving `NaN` produce inconsistent results between cudf.polars and polars.

Polars seems to treat NaN as the largest value in comparisons, whereas cudf.polars follows IEEE 754 (as discussed in https://github.com/rapidsai/cudf/issues/4760). I'm not sure whether this difference is a bug or intended behavior

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

lf = pl.LazyFrame(
{
"a": [None,1., np.nan],
}
)

print(lf.select(pl.col('a')>2).collect())

print(lf.select(pl.col('a')>2).collect(engine='gpu'))

"""
shape: (3, 1)
┌───────┐
│ a │
│ --- │
│ bool │
╞═══════╡
│ null │
│ false │
│ true │
└───────┘

shape: (3, 1)
┌───────┐
│ a │
│ --- │
│ bool │
╞═══════╡
│ null │
│ false │
│ false │
└───────┘
"""
```