Simplify GPU quantile tests and tighten sketch container ownership

[RAMitchell]

Summary

This continues the GPU quantile test rewrite and tightens the GPU SketchContainer
interface so the sketch owns more of its own sizing and maintenance behavior.

What Changed

• simplify the low-level GPU sketch tests in tests/cpp/common/test_quantile.cu
  • use direct SketchContainer::Push(...) input for push/prune/merge coverage
  • remove tests that belonged at the wrapper/cut layer instead of the sketch layer
  • keep explicit duplicate/rank coverage for Push(...)
• align GPU property-style cut tests with the shared CPU helper layer
  • shared container cases
  • shared cut invariant validation
• tighten the GPU sketch container interface
  • SketchContainer::Push(...) now computes its own batch-local cut layout
  • internal maintenance prune uses a per-feature bound
  • remove external post-batch prune from QuantileDMatrix
• remove redundant GPU wrapper tests from tests/cpp/common/test_hist_util.cu
• simplify categorical dedup bookkeeping by updating the new column scan buffer directly

Testing

Built testxgboost with CUDA and ran focused GPU coverage, including:

• GPUQuantile.*
• GPUQuantileProperty.Invariants
• MGPUQuantileTest.SameOnAllWorkers
• focused remaining HistUtil GPU tests

Local result:

• focused GPU low-level slice passed
• SameOnAllWorkers still skips locally without federated/NCCL support

[Copilot]

Pull request overview

This PR refactors GPU quantile sketch tests and tightens the GPU SketchContainer interface so the container computes its own per-batch cut layout and owns more of its maintenance behavior.

Changes:

• Simplify/realign GPU quantile tests to exercise SketchContainer::Push/Prune/Merge/AllReduce more directly and share CPU/GPU cut invariant validation via helpers.
• Update GPU sketch container API so Push(...) computes batch-local cut pointers internally; remove redundant external prune in QuantileDMatrix.
• Simplify categorical dedup bookkeeping by updating the column scan buffer directly; remove redundant GPU wrapper tests from test_hist_util.cu.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/cpp/common/test_quantile.h	Updates helper header guard and seed generation used by quantile tests.
tests/cpp/common/test_quantile.cu	Major GPU quantile test rewrite: new synthetic batch helpers + property-style tests; updated calls to new SketchContainer::Push signature.
tests/cpp/common/test_quantile.cc	Replaces local container-invariant logic with shared ValidateContainerCuts helper.
tests/cpp/common/test_quantile_helpers.h	Adds shared ValidateContainerCuts(...) helper used by both CPU/GPU tests.
tests/cpp/common/test_hist_util.cu	Removes redundant DeviceSketch wrapper tests; updates categorical dedup test for new API.
src/data/quantile_dmatrix.cu	Removes per-batch external prune now handled internally by the sketch container.
src/common/quantile.cuh	Tightens SketchContainer::Push signature; renames intermediate sizing helper to per-feature semantics.
src/common/quantile.cu	Implements internal batch-local cut pointer construction and prunes using a per-feature bound.
src/common/hist_util.cuh	Removes external cut-pointer plumbing; relies on SketchContainer::Push to compute cut layout.
src/common/hist_util.cu	Simplifies categorical dedup to update column scan only; updates weighted sketch path to new Push API.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 11 out of 11 changed files in this pull request and generated 4 comments.

Comments suppressed due to low confidence (1)

src/common/quantile.cuh:157

• SketchContainer::Prune(Context const* ctx, size_t to) treats to as a per-feature cap (it does length = std::min(length, to) for each column). The current docstring says "maximum size of pruned quantile" which reads like a global/total cap. Updating the comment to explicitly say "maximum entries per feature" would make the API contract clearer, especially now that Push() prunes internally using IntermediateCutsPerFeature().

  /**
   * @brief Prune the quantile structure.
   *
   * @param to The maximum size of pruned quantile.  If the size of quantile structure is
   *           already less than `to`, then no operation is performed.
   */
  void Prune(Context const* ctx, size_t to);

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.