fix(cudf): Restore Spark workload coverage

[thirtiseven]

Summary

• restore cuDF evaluation for Spark regexp_extract, numeric-to-VARCHAR
  casts, isnull, dynamic array, and multi-branch switch expressions
• restore CudfUnnest to the cuDF exec build and operator adapter registry
• allow partial-identity aggregation without a streaming distinct-key capacity
• transfer partial-identity input column ownership instead of deep-copying every grouping key and aggregate input

Motivation

The TypedExpr expression-subsystem merge dropped coverage that existed in the previous exec::Expr implementation. Strict functional canaries exposed regressions across scalar expression evaluation, multi-branch switch, and unnest operator selection.

Large-scale aggregation validation additionally exposed two independent partial-identity issues: enabling partial identity was incorrectly coupled to a non-zero streaming capacity, and partial identity deep-copied its complete input. The ownership transfer follows the existing CudfFilterProject pattern and copies only when a source column has multiple consumers.

Test plan

Native GPU tests:

• AggregationTest.partialIdentityUsesQueryScopedStreamingCapacity
• AggregationTest.partialIdentityDoesNotRequireStreamingCapacity
• CudfExpressionSelectionTest.multiBranchSwitch

Production build:

• velox_cudf_exec compiled successfully
• Spark Gluten 4.0 / Java 17 / Scala 2.13 bundle packaged successfully

100% local MPP validation (4 executors / 4 x 32 GiB GPUs, CPU fallback disabled):

• unnest-heavy query: SUCCESS, 95.743 s with the normal async allocator
• aggregation-heavy query: SUCCESS, 465.777 s with managed_async

The aggregation-heavy query still outgrows a 32 GiB GPU with async. Device diagnostics show about 33.25 GB of live RMM allocations while the FINAL group-by accumulates state, so this is not retained allocator cache. managed_async validates the fixed operator paths and correctness at full scale, but it is not a performance configuration. A separate follow-up should add FINAL group-by spill or test a topology with more physical MPP partitions.

[Copilot]

Pull request overview

Restores cuDF coverage for Spark expressions, unnest execution, and partial-identity aggregation.

Changes:

• Adds Spark expression support and parity tests.
• Re-registers CudfUnnest.
• Optimizes partial-identity aggregation ownership and capacity handling.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
FilterProjectTest.cpp	Adds expression parity tests.
ExpressionEvaluatorSelectionTest.cpp	Tests GPU expression selection.
AggregationTest.cpp	Tests zero-capacity partial identity.
AdapterOperatorTest.cpp	Verifies cuDF unnest selection.
ExpressionEvaluator.cpp	Implements restored Spark expressions.
AstExpressionUtils.h	Adds the isnull alias.
OperatorAdapters.cpp	Registers the unnest adapter.
CudfGroupby.h	Extends partial-identity ownership API.
CudfGroupby.cpp	Transfers input ownership during partial identity.
CMakeLists.txt	Restores CudfUnnest.cpp to the build.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[winningsix]

there's a failing CI job. Please help address it before merging.

[thirtiseven]

there's a failing CI job. Please help address it before merging.

It is failing for a while and some changes is not related to this pr. Let me try to fix it in another PR.