[SPARK-58067][PYTHON][TESTS] Add ASV microbenchmark for SQL_TRANSFORM_WITH_STATE_PYTHON_ROW_INIT_STATE_UDF

[Yicong-Huang]

What changes were proposed in this pull request?

Add an ASV microbenchmark for the SQL_TRANSFORM_WITH_STATE_PYTHON_ROW_INIT_STATE_UDF eval type (served by TransformWithStateInPySparkRowInitStateSerializer), mirroring the existing SQL_TRANSFORM_WITH_STATE_PANDAS_INIT_STATE_UDF benchmark and the non-init Row benchmark (#57047). It reuses the existing bench_eval_type.py harness and the plain-Row scenario grid from _TransformWithStateRowBenchMixin, adding TransformWithStateRowInitStateUDF{Time,Peakmem}Bench over 7 scenarios x 3 UDFs. The UDF signature is (api_client, mode, key, rows, init_rows) where both rows and init_rows are iterators of Row. The input wire stream is the nested struct<inputData, initState> Arrow stream (init batches first, then data batches); the serializer materializes every column into a Row via .as_py() and regroups by the leading key, so each key surfaces as one init-only call followed by one data-only call -- the per-row Python object path this eval type is built around, layered on the init-state deserialization.

Stacked on #57047 (_TransformWithStateRowBenchMixin); will rebase onto master once that merges.

Why are the changes needed?

Establishes a no-regression baseline for the upcoming refactor that moves the Row init-state transformWithState logic out of the serializer into read_udfs().

Does this PR introduce any user-facing change?

No.

How was this patch tested?

Benchmark-only change. Ran asv run --bench 'TransformWithStateRowInitState' -a repeat=3 --python=same twice; results were stable across runs (time within ~2%, peakmem identical). One representative run:

[time_worker]
================ ============== ============= ============
--                                  udf
---------------- -----------------------------------------
    scenario      identity_udf   rebuild_udf   count_udf
================ ============== ============= ============
 few_groups_sm      475±3ms        529±2ms      391±1ms
 few_groups_lg     4.04±0.02s     4.60±0.01s    3.24±0s
 many_groups_sm    4.73±0.02s     4.96±0.02s   4.20±0.01s
 many_groups_lg    2.39±0.01s     2.62±0.03s    2.01±0s
   wide_cols       5.14±0.01s     5.62±0.02s    4.24±0s
   mixed_cols      1.96±0.01s     2.19±0.01s   1.62±0.01s
 nested_struct      3.31±0s       3.51±0.01s    2.68±0s
================ ============== ============= ============

[peakmem_worker]
================ ============== ============= ===========
    scenario      identity_udf   rebuild_udf   count_udf
================ ============== ============= ===========
 few_groups_sm       92.2M          92.2M        88.7M
 few_groups_lg        114M           114M         104M
 many_groups_sm       105M           105M         102M
 many_groups_lg      98.7M          98.7M        92.6M
   wide_cols          113M           113M         113M
   mixed_cols         101M           101M        97.1M
 nested_struct        103M           103M         103M
================ ============== ============= ===========

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

No.