[BUG] Investigate Narwhals tests failures when run with cudf.pandas

[Matt711]

Describe the bug
In #18031, we started running the Narwhals test suite with cudf.pandas. There are several failing tests that we needed to skip (listed in https://github.com/rapidsai/cudf/blob/branch-25.04/ci/test_narwhals.sh). We should either fix these failures or explain why they won't be fixed.

Steps/Code to reproduce bug
Follow the steps in this shell scripts (eg. clone Narwhals, install dependencies, etc.).

Additional context
Please xref this issue once you've fixed a test

Original Issue to add Narwhals tests to cuDF: #17662

[galipremsagar]

I'm investigating the cross join test failures.

[mroeschke]

#18255 will help partially resolve the skipped test_pandas_object_series test, but I also think Narwhals would need to add an explicit dtype=object call here in https://github.com/narwhals-dev/narwhals/blob/af20c95c1e35e3dd4183b7d001ea69fd142d718b/tests/expr_and_series/list/len_test.py#L66

since cudf infers that data as ListDtype and pandas infers that data as object by default

[mroeschke]

test_array_dunder_with_copy[pandas] I think should always be skipped. It's testing numpy copy behavior with __array__, and we always make a copy to CPU here

[mroeschke]

tests/utils_test.py::test_maybe_convert_dtypes_pandas should fail until #14149 is addressed

[mroeschke]

tests/series_only/to_arrow_test.py::test_to_arrow_with_nulls[pandas] and tests/frame/to_arrow_test.py::test_to_arrow[pandas] look OK to skip.

They are testing a Narwhals implementation of to_arrow which looks to have different behavior whether cudf is involved or not https://github.com/narwhals-dev/narwhals/blob/af20c95c1e35e3dd4183b7d001ea69fd142d718b/narwhals/_pandas_like/series.py#L879. This test doesn't really consider if cudf.pandas is loaded (nor should really care)