[BUG]: Bug in cudf-polars to_string for datetime column with timezone

[TomAugspurger]

Describe the bug

cudf-polars doesn't match polars to_string when the datetime column has a timezone.

def test_to_string(engine: pl.GPUEngine) -> None:
    data = {
        "a": [
            datetime.datetime(2020, 1, 1, tzinfo=datetime.UTC),
            datetime.datetime(2020, 1, 2, tzinfo=datetime.UTC),
        ]
    }
    df = pl.LazyFrame(data)
    q = df.select(pl.col("a").dt.replace_time_zone(None).dt.convert_time_zone("Asia/Kathmandu"))
    assert_gpu_result_equal(q, engine=engine)

    q2 = q.select(pl.col("a").dt.to_string("%Y-%m-%dT%H:%M%z"))
    assert_gpu_result_equal(q2, engine=engine)

with

AssertionError: DataFrames are different (value mismatch for column "a")
[left]: shape: (2,)
Series: 'a' [str]
[
        "2020-01-01T05:45+0545"
        "2020-01-02T05:45+0545"
]
[right]: shape: (2,)
Series: 'a' [str]
[
        "2020-01-01T00:00+0000"
        "2020-01-02T00:00+0000"
]
=========================================================================================================================================================== short test summary info ============================================================================================================================================================
FAILED python/cudf_polars/tests/expressions/test_datetime_timezone.py::test_to_string[in-memory] - AssertionError: DataFrames are different (value mismatch for column "a")

This is causing CI failures in the narhwals tests: https://github.com/NVIDIA/cudf/actions/runs/33631705030/job/100262514357?pr=23922#step:14:4243