[feat][spark-4.0.2] Query optimization / AQE changes (runtime filter, AQE InMemoryTableScan, column order, clustering, SPJ) on GPU (cuDF)

[winningsix]

Background

Velox-native (cuDF) counterpart of the NVIDIA/spark-gluten issue for this Spark 4.0.2 feature. Depends on the native harness #66 and the JVM harness NVIDIA/spark-gluten#134.
Spark tickets:

• SPARK-45606
• SPARK-45592
• SPARK-50258
• SPARK-46693
• SPARK-45784
• SPARK-51938

Scope (native cuDF side)

• Implement / extend the cuDF GPU operators and expressions under velox/experimental/cudf/ backing this feature.
• Add / extend gtests under velox/experimental/cudf/tests/.
• Ensure CPU Velox tests are unaffected (VELOX_ENABLE_CUDF-gated).

Acceptance criteria — all pass on GPU (VELOX_ENABLE_CUDF=ON, ctest)

• HashJoinTest
• OrderByTest
• TopNTest

Cross-references

• Spark-Gluten (JVM) counterpart: NVIDIA/spark-gluten#155 (https://github.com/NVIDIA/spark-gluten/issues/155).
• Harness: [feat][spark-4.0.2] cuDF native test harness: build & run velox/experimental/cudf/tests on Spark-Flux GPU CI #66, NVIDIA/spark-gluten#134.

Notes

Multi-layer runtime filter; AQE InMemoryTableScan correctness; Storage Partition Join.