[SPARK-57423][PYTHON][TEST] Share base mixin across MapInBatch Arrow/Pandas siblings

[viirya]

What changes were proposed in this pull request?

Follow-up to SPARK-57137/57138/57414. Apply the sibling base-mixin pattern to the last remaining Arrow/Pandas pair in python/benchmarks/bench_eval_type.py:

• _MapPandasIterBenchMixin now subclasses _MapArrowIterBenchMixin, dropping its duplicated _build_scenario (identical body) and _write_scenario (identical except the eval type constant and the ret_type handling).
• The Arrow base gains the _eval_type class attribute and parameterizes _write_scenario on self._eval_type. The Pandas _udfs values are normalized from (func, arg_offsets) to the Arrow base's (func, ret_type, arg_offsets) shape so the inherited writer works unchanged.
• Drive-by: convert the _build_scenario of the newly added _TransformWithStatePandasBenchMixin (SPARK-57020, developed in parallel with SPARK-57414) from @staticmethod with a hardcoded class-name lookup to @classmethod with cls._scenario_configs, matching the now-uniform convention.

Net diff: +16 / -40 lines.

Why are the changes needed?

The two MapInBatch mixins were the last undeclared sibling pair: near-identical _build_scenario and _write_scenario bodies that had to be kept in lock-step manually. After this change, every Arrow/Pandas sibling pair in the file shares a base mixin, every _build_scenario resolves configs via cls, and no redundant protocol-writing copies remain.

Does this PR introduce any user-facing change?

No. Test-only change in the benchmark module.

How was this patch tested?

• Verified the generated worker input is byte-identical to the pre-refactor output for all 21 *TimeBench classes across every (scenario, udf) cell, with the cloudpickled UDF command excluded and the TWS stub state-server port pinned (the port is freshly allocated per process; the UDF pickle embeds co_firstlineno and a random cloudpickle class-tracker id -- all environment/location metadata unaffected by this change).
• Compared the pickle opcode streams of all 63 pickled UDF/UDTF callables between old and new: only the metadata above differs; code, constants, and names are unchanged.
• Ran setup + time_worker end-to-end for the affected *TimeBench classes (MapArrowIter, MapPandasIter, TransformWithStatePandas) across every UDF, and peakmem_worker for MapPandasIter and TransformWithStatePandas.

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

Yes. Generated-by: Claude Code (claude-opus-4-8)