What changes were proposed in this pull request?

Refactor SQL_GROUPED_AGG_PANDAS_UDF to use ArrowStreamGroupSerializer as a pure I/O layer, moving the per-group pandas conversion and UDF invocation into read_udfs() in worker.py. The custom ArrowStreamAggPandasUDFSerializer is no longer used for this eval type (still used by SQL_GROUPED_AGG_PANDAS_ITER_UDF and SQL_WINDOW_AGG_PANDAS_UDF).

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. No behavior change.

ASV benchmark comparison (GroupedAggPandasUDFTimeBench, 3 samples per scenario, min taken):

Scenario         UDF                Before     After   Change
few_groups_sm    sum_udf            33.46ms   27.31ms  -18.4%
few_groups_sm    mean_multi_udf     35.42ms   29.24ms  -17.4%
few_groups_lg    sum_udf            52.52ms   41.37ms  -21.2%
few_groups_lg    mean_multi_udf     58.26ms   46.82ms  -19.6%
many_groups_sm   sum_udf          1223.22ms 1012.82ms  -17.2%
many_groups_sm   mean_multi_udf   1295.45ms 1059.40ms  -18.2%
many_groups_lg   sum_udf           350.11ms  285.29ms  -18.5%
many_groups_lg   mean_multi_udf    377.48ms  309.53ms  -18.0%
wide_cols        sum_udf           362.97ms  286.47ms  -21.1%
wide_cols        mean_multi_udf    378.80ms  299.86ms  -20.8%

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

No.