[mt] Feature importance variants.

[trivialfis]

• Gain and coverage.

At the moment, I store the sum hessian for all targets. We might need to store a vector instead. I will leave it as future work if it comes up.

Ref #9043

todos:

• Test mixed tree types.

[Copilot]

Pull request overview

Adds multi-target tree support for additional feature importance variants (gain/cover and totals) by storing and exposing per-node statistics, plus expands test coverage across CPU/GPU and mixed multi-strategy training.

Changes:

• Persist and expose loss_chg (gain) and sum_hess (cover) for multi-target trees, including JSON serialization.
• Update CPU/GPU multi-target training paths and tree views to populate/read these statistics efficiently (optionally skipping stats for inference paths).
• Add/relocate tests for feature importance variants, column sampling, and mixed multi-strategy behavior on CPU and GPU.

Reviewed changes

Copilot reviewed 21 out of 21 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
tests/python/test_multi_target.py	Adds CPU test entrypoints for new multi-target test helpers.
tests/python/test_basic_models.py	Removes multi-tree slicing test (coverage moved to multi-target suite).
tests/python-gpu/test_gpu_multi_target.py	Adds GPU tests for mixed strategy and feature-importance comparison.
tests/cpp/tree/test_partitioner.h	Updates test helpers to new SetRoot/ExpandNode multi-target signatures.
tests/cpp/tree/test_multi_target_tree_model.cc	Updates signatures and adds serialization round-trip test for stats.
tests/cpp/tree/hist/test_evaluate_splits.cc	Ensures root sum_hess is set for multi-target hist evaluator tests.
tests/cpp/predictor/test_predictor.cc	Updates predictor test tree construction to include multi-target stats.
src/tree/updater_quantile_hist.cc	Computes and sets root sum_hess for multi-target quantile hist builder.
src/tree/updater_gpu_hist.cuh	Populates multi-target node stats on GPU hist path; avoids stats for partitioning.
src/tree/tree_view.h	Adds multi-target stat pointers and new constructor flag need_stat.
src/tree/tree_view.cc	Implements conditional stat loading for multi-target tree views.
src/tree/tree_model.cc	Updates multi-target ExpandNode signature plumbing.
src/tree/multi_target_tree_model.cc	Stores stats in MultiTargetTree, and saves/loads them in JSON.
src/tree/hist/evaluate_splits.h	Populates multi-target split stats (gain/cover) on CPU hist expand.
src/tree/gpu_hist/multi_evaluate_splits.cu	Tracks summed child hessians across all targets for GPU multi-evaluator.
src/tree/gpu_hist/expand_entry.cuh	Renames/repurposes hessian fields to summed left/right hessians.
src/predictor/gbtree_view.h	Passes need_stat through when building per-tree views.
src/gbm/gbtree.h	Enables gain/cover feature importance for multi-target trees using new stats.
python-package/xgboost/testing/multi_target.py	Adds new Python test helpers covering importance variants and strategies.
include/xgboost/tree_model.h	Extends multi-target SetRoot/ExpandNode APIs with stat parameters.
include/xgboost/multi_target_tree_model.h	Extends multi-target tree structure to store loss change and hessian sums.

---

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

[trivialfis]

We can have vector leaf for #11926 as well

[trivialfis]

cc @rongou