[SPARK-56123][PYTHON] Refactor SQL_GROUPED_AGG_ARROW_UDF and SQL_GROUPED_AGG_ARROW_ITER_UDF

[Yicong-Huang]

What changes were proposed in this pull request?

Refactor SQL_GROUPED_AGG_ARROW_UDF and SQL_GROUPED_AGG_ARROW_ITER_UDF to use ArrowStreamSerializer as a pure I/O layer, moving all processing logic into read_udfs() in worker.py.

Why are the changes needed?

Part of SPARK-55388.

Does this PR introduce any user-facing change?

No.

How was this patch tested?

Existing tests.

ASV micro-benchmarks with repeat=(3, 5, 5.0) show consistent improvements:

SQL_GROUPED_AGG_ARROW_UDF (time)

Scenario	UDF	Before	After	Change
few_groups_sm	sum	8.32ms	7.28ms	-12%
few_groups_sm	mean_multi	7.38ms	6.01ms	-19%
few_groups_lg	sum	29.7ms	29.7ms	0%
few_groups_lg	mean_multi	30.4ms	29.9ms	-2%
many_groups_sm	sum	231ms	181ms	-22%
many_groups_sm	mean_multi	206ms	144ms	-30%
many_groups_lg	sum	109ms	95.9ms	-12%
many_groups_lg	mean_multi	107ms	85.0ms	-21%
wide_cols	sum	75.9ms	59.9ms	-21%
wide_cols	mean_multi	74.7ms	57.1ms	-24%

SQL_GROUPED_AGG_ARROW_ITER_UDF (time)

Scenario	UDF	Before	After	Change
few_groups_sm	sum	6.41ms	5.66ms	-12%
few_groups_sm	mean_multi	5.37ms	4.63ms	-14%
few_groups_lg	sum	18.2ms	17.9ms	-2%
few_groups_lg	mean_multi	20.7ms	20.3ms	-2%
many_groups_sm	sum	190ms	166ms	-13%
many_groups_sm	mean_multi	156ms	131ms	-16%
many_groups_lg	sum	78.4ms	72.8ms	-7%
many_groups_lg	mean_multi	71.1ms	65.5ms	-8%
wide_cols	sum	49.1ms	44.0ms	-10%
wide_cols	mean_multi	45.7ms	39.8ms	-13%

Peak memory: No change (~1.15G for all scenarios).

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

No.

[Yicong-Huang]

This PR depends on #54967 (enforce_schema). Will rebase after that one merges.

[zhengruifeng]

merged to master