[SPARK-59443][SQL] Reduce allocations in ArrowColumnVector decimal reads

[xiaoh1024]

What changes were proposed in this pull request?

Add a small-decimal accessor for Arrow columns with source precision 1-18,
with full-width fallback and existing Spark decimal conversion semantics.
Add regression tests and an Arrow decimal read benchmark.

Why are the changes needed?

See SPARK-59443.

Does this PR introduce any user-facing change?

No.

How was this patch tested?

• build/sbt catalyst/compile sql/Test/compile
• build/sbt "sql/testOnly org.apache.spark.sql.vectorized.ArrowColumnVectorSuite"
  (28 tests passed, including 8 new decimal tests).
• The same suite passed all 28 tests on a big-endian s390x Temurin 17 JVM
  under QEMU, including a native Arrow Decimal128 layout check.
• Checkstyle for the changed Java file and Scalastyle for both Scala files.
• Local ArrowDecimalReadBenchmark and allocation comparisons against the
  original accessor.
• ArrowDecimalReadBenchmark on GitHub Actions
  passed with JDK 17 and Scala 2.13.
• Fork Build CI passed.

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

AI assistance was used for implementation, tests, and local validation.

Generated-by: Codex CLI 0.154.0

[sunchao]

Reviewed 0710e00. No actionable correctness findings.

Traced accessor selection, Arrow 19's native Decimal128 layout, the signed-long/full-width fallback, and Spark's BigDecimal-backed rounding, overflow, and integral-conversion semantics. Also reviewed the regression tests and benchmark baseline.

Independent validation: a standalone JDK 17 harness using the two accessor classes extracted unchanged from this PR, the PR's compiled Decimal.scala, Arrow 19.0.0, and cached Spark 4.2 supporting jars passed 63,523 old/new accessor comparisons. Coverage included 20,736 randomized rows, values beyond the declared precision and signed-long range, requested precision/scale changes, nulls, buffer reallocation, slicing, and independent returned objects. A separate BigDecimal-constructor parity check also passed.

The fork's build/test checks are successful at this commit. The local harness is focused validation; I did not rerun the full Spark ArrowColumnVectorSuite, the performance benchmark, or native big-endian execution.

[sunchao]

Approved based on the completed review and focused validation at this unchanged commit.