window_lead test appears to be non-deterministic

[andygrove]

Describe the bug

The test works for me locally but fails in CI/

2023-01-19T00:06:25.0267445Z df = <datafusion.DataFrame object at 0x7ffbd80ccc70>
2023-01-19T00:06:25.0268026Z 
2023-01-19T00:06:25.0268336Z     def test_window_lead(df):
2023-01-19T00:06:25.0268610Z         df = df.select(
2023-01-19T00:06:25.0268869Z             column("a"),
2023-01-19T00:06:25.0269119Z             f.alias(
2023-01-19T00:06:25.0269347Z                 f.window(
2023-01-19T00:06:25.0269666Z                     "lead", [column("b")], order_by=[f.order_by(column("b"))]
2023-01-19T00:06:25.0269979Z                 ),
2023-01-19T00:06:25.0270215Z                 "a_next",
2023-01-19T00:06:25.0270448Z             ),
2023-01-19T00:06:25.0270671Z         )
2023-01-19T00:06:25.0270865Z     
2023-01-19T00:06:25.0271152Z         table = pa.Table.from_batches(df.collect())
2023-01-19T00:06:25.0271426Z     
2023-01-19T00:06:25.0271707Z         expected = {"a": [1, 2, 3], "a_next": [5, 6, None]}
2023-01-19T00:06:25.0272031Z >       assert table.to_pydict() == expected
2023-01-19T00:06:25.0272883Z E       AssertionError: assert {'a': [3, 1, ... [None, 5, 6]} == {'a': [1, 2, ... [5, 6, None]}
2023-01-19T00:06:25.0273211Z E         Differing items:
2023-01-19T00:06:25.0273583Z E         {'a_next': [None, 5, 6]} != {'a_next': [5, 6, None]}
2023-01-19T00:06:25.0273949Z E         {'a': [3, 1, 2]} != {'a': [1, 2, 3]}
2023-01-19T00:06:25.0274208Z E         Full diff:
2023-01-19T00:06:25.0274555Z E         - {'a': [1, 2, 3], 'a_next': [5, 6, None]}
2023-01-19T00:06:25.0274917Z E         ?            ---              ------
2023-01-19T00:06:25.0275260Z E         + {'a': [3, 1, 2], 'a_next': [None, 5, 6]}
2023-01-19T00:06:25.0275725Z E         ?        +++                      ++++++

To Reproduce
Steps to reproduce the behavior:

Expected behavior
A clear and concise description of what you expected to happen.

Additional context
Add any other context about the problem here.