[SPARK-56424][PYTHON] Add ASV benchmark for SQL_SCALAR_PANDAS_UDF

[Yicong-Huang]

What changes were proposed in this pull request?

Add ASV microbenchmarks for SQL_SCALAR_PANDAS_UDF eval type in bench_eval_type.py.

The new _ScalarPandasBenchMixin follows the same pattern as the existing _ScalarArrowBenchMixin but uses Pandas Series operations. It measures the full Arrow-to-Pandas-to-Arrow round-trip that occurs in scalar Pandas UDFs.

Scenarios (9): sm_batch_few_col, sm_batch_many_col, lg_batch_few_col, lg_batch_many_col, pure_ints, pure_floats, pure_strings, pure_ts, mixed_types

UDFs (3): identity_udf (passthrough), sort_udf (Series.sort_values), nullcheck_udf (Series.notna)

Benchmark classes (2): ScalarPandasUDFTimeBench, ScalarPandasUDFPeakmemBench

Why are the changes needed?

This is part of the PySpark Serializer & EvalType Refactor effort (SPARK-55724). We need baseline benchmarks for every eval type before refactoring the serialization path, so we can detect performance regressions. SQL_SCALAR_PANDAS_UDF is one of the most commonly used eval types and its Arrow-to-Pandas-to-Arrow conversion cost is a key metric for the refactor.

Does this PR introduce any user-facing change?

No.

How was this patch tested?

Ran all 54 benchmarks (27 time + 27 peakmem) locally with python/asv run --python=same --quick:

time_worker results (27/27 passed):
  sm_batch_few_col   identity_udf: 269ms, sort_udf: 385ms, nullcheck_udf: 300ms
  lg_batch_few_col   identity_udf: 836ms, sort_udf: 1.06s, nullcheck_udf: 834ms
  pure_ints          identity_udf: 165ms, sort_udf: 236ms, nullcheck_udf: 170ms
  pure_strings       identity_udf: 819ms, sort_udf: 1.17s, nullcheck_udf: 818ms
  mixed_types        identity_udf: 504ms, sort_udf: 551ms, nullcheck_udf: 510ms
  (+ 12 more scenarios all passing)

peakmem_worker results (27/27 passed):
  Range: 479M - 628M across all scenario/udf combinations

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

No.