[mt] Unify the quantisers.

[trivialfis]

• Support distributed training.
• Host device access. Host access is for single-target. We will migrate to device-only once we can unify the code base between single-target and multi-target more.

ref #9043

[Copilot]

Pull request overview

This PR unifies the gradient quantiser implementation to support distributed training for multi-output models. The changes introduce a new GradientQuantiserGroup class that replaces the previous MultiGradientQuantiser and provides both device-side access (via DeviceSpan()) and host-side access (via operator[]) to quantisers. The implementation moves quantiser construction logic into a new quantiser.cu file and optimizes distributed training by batching multiple collective operations together.

Changes:

• Introduced unified GradientQuantiserGroup class replacing MultiGradientQuantiser with dual host/device access
• Moved quantiser implementation to new quantiser.cu file with optimized collective operations for distributed training
• Updated all quantiser access patterns from .Quantizers() to .DeviceSpan() for device access and operator[] for host access
• Modified distributed training check to allow distributed training for non-vector-leaf cases

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/tree/gpu_hist/quantiser.cuh	Defines new unified GradientQuantiserGroup class with host/device dual interface
src/tree/gpu_hist/quantiser.cu	New file containing quantiser implementation with batched collective operations for distributed training
src/tree/gpu_hist/histogram.cu	Removed quantiser implementation code (moved to quantiser.cu)
src/tree/updater_gpu_hist.cu	Updated single-target quantiser usage to use new GradientQuantiserGroup with operator[] for host access
src/tree/updater_gpu_hist.cuh	Updated multi-target quantiser usage and relaxed distributed training restriction to allow non-vector-leaf cases
tests/cpp/tree/gpu_hist/test_sampler.cu	Updated test to use new DeviceSpan() API
tests/cpp/tree/gpu_hist/test_histogram.cu	Updated tests to use GradientQuantiserGroup with both host and device access patterns
tests/cpp/tree/gpu_hist/test_evaluate_splits.cu	Updated test helper to construct quantiser via GradientQuantiserGroup
tests/cpp/tree/gpu_hist/dummy_quantizer.cuh	Updated test utilities to use GradientQuantiserGroup
src/predictor/interpretability/shap.cu	Minor cleanup to suppress unused parameter warning

Comments suppressed due to low confidence (1)

src/tree/gpu_hist/quantiser.cu:51

• The LaTeX equation syntax in the documentation comment uses @begin{equation} and @end{equation}, which is not standard Doxygen syntax. Doxygen typically uses \f[ and \f] for display equations or \f$ for inline equations. Consider changing to \f[ and \f] to ensure the equation renders correctly in generated documentation.

 * @begin{equation}
 *   max( fl(\sum^{V}_{v_i>0}{v_i}), fl(\sum^{V}_{v_i<0}|v_i|) )
 * @end{equation}

---

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

[trivialfis]

cc @rongou