[BUG] Root cause and fix flaky failures in Polars CI runs due to the rapidsmpf transition

Since we switched over to using the rapidsmpf runtime in our GPU polars CI runs we've seen some tests fail intermittently in CI. I skipped a couple of them in https://github.com/rapidsai/cudf/pull/22294, but we've seen additional failures as well:
- operations/test_join.py::test_join_filter_pushdown_cross_join in [this run](https://github.com/rapidsai/cudf/actions/runs/24980827101/job/73142411357)
- operations/test_join.py::test_join_coalesce_not_supported_warning in [this run](https://github.com/rapidsai/cudf/actions/runs/24950563573/job/73059814023)

A couple of the failures are `AssertionError: DataFrames are different (height (row count) mismatch)` but that's not the only failure.