[SPARK-58058][PYTHON][TESTS] Split slow PySpark test suites into smaller modules

[HyukjinKwon]

What changes were proposed in this pull request?

This PR splits several large/slow PySpark test suites into smaller modules so
they run as separate test targets and parallelize better, reducing wall-clock
time. No test logic is changed; test methods are only relocated, following the
existing convention already used in the repo (e.g. test_split_apply_*,
test_stat_*).

The following suites are split, each with a matching Spark Connect parity module:

• pyspark.sql.tests.pandas.test_pandas_cogrouped_map: move the heavier,
  self-contained tests (test_with_window_function, test_with_local_data,
  test_arrow_batch_slicing, test_cogroup_apply_in_pandas_with_logging) into
  a new test_pandas_cogrouped_map_misc module.
• pyspark.sql.tests.arrow.test_arrow_cogrouped_map: extract shared helpers
  into CogroupedMapInArrowTestsFuncMixin and move the heavier tests into a new
  test_arrow_cogrouped_map_misc module.
• pyspark.pandas.tests.data_type_ops.test_num_ops: move
  IntegralExtensionOpsTestsMixin and FractionalExtensionOpsTestsMixin into
  their own modules.
• pyspark.pandas.tests.groupby.test_stat: move test_median into a new
  test_stat_median module.
• pyspark.pandas.tests.groupby.test_split_apply: move the mean case into a
  new test_split_apply_mean module.

New modules are registered in dev/sparktestsupport/modules.py.

Why are the changes needed?

These suites are slow in CI. Splitting them lets the test runner schedule the
pieces in parallel, lowering total wall-clock time without dropping coverage.

Does this PR introduce any user-facing change?

No. Test-only change.

How was this patch tested?

Existing tests, only relocated across modules. Verified that the total number
of test methods is unchanged for each split suite, that all new and modified
modules import and collect, and that flake8 passes.

Was this patch authored or co-authored using generative AI tooling?

Generated-by: Cursor

[HyukjinKwon]

CI timing results (run on this PR)

All split modules pass. The one red job (pyspark-sql, pyspark-core, ...) failed only on pyspark.tests.test_import_spark (Unexpected 3rd party package 'numpy' imported during 'import pyspark'), which is unrelated to this test-only change and not part of this diff.

Measured per-module durations (classic / Spark Connect):

Suite	base	new module(s)
arrow cogrouped map	22s	misc 23s
pandas cogrouped map	41s	misc 25s
num_ops	27s / 28s	integral_ext 37s / 50s, fractional_ext 26s / 40s
groupby stat	27s / 36s	stat_median 12s / 15s
split_apply	29s / 41s	split_apply_mean 29s / 41s

Each split lowers the longest single-module wall-clock via parallelism. arrow and split_apply are well balanced; num_ops, groupby stat, and pandas cogrouped map are functional splits but not tightly balanced (documented here for transparency; can be rebalanced in a follow-up if desired).

[dongjoon-hyun]

I also agree that pyspark.tests.test_import_spark failure is irrelevant.

[HyukjinKwon]

Merge Summary:

• merged into master 79f6dfa
• merged into branch-4.x 89b66fe

Posted by merge_spark_pr.py