Move DART state into GBTree

[RAMitchell]

This PR is the next step toward removing GBDart as a separate booster. It moves the DART runtime state and training-time dropout behavior into GBTree, so dropout can be enabled directly through parameters like rate_drop without requiring booster=dart. The legacy Dart class remains as a thin compatibility wrapper for the existing booster name and serialized model/config shape. A follow-up PR will load old DART models/configs directly into unified GBTree state and remove the separate Dart class.

[Copilot]

Pull request overview

This PR moves DART (Dropout Additive Regression Trees) runtime state and training-time dropout behavior from the separate Dart class into the base GBTree class. This allows dropout to be enabled directly through parameters like rate_drop when using booster=gbtree, without requiring booster=dart. The Dart class is retained as a thin compatibility wrapper for existing serialized models/configs.

Changes:

• Moved weight_drop_, idx_drop_, dparam_, DropTrees(), NormalizeTrees(), and dropout-aware prediction logic from Dart into GBTree.
• Simplified Dart to a thin wrapper that only overrides serialization methods (SaveModel, LoadModel, SaveConfig, LoadConfig) and CommitModel for backward compatibility with the "dart" JSON format.
• Updated documentation (parameter.rst, dart.rst) and tests to reflect that dropout is now a feature of gbtree rather than a separate booster.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/gbm/gbtree.h	Added dparam_, weight_drop_, idx_drop_ members and DropTrees/NormalizeTrees declarations to GBTree; made TreeWeights non-virtual.
src/gbm/gbtree.cc	Moved dropout logic (DropTrees, NormalizeTrees, weight tracking in CommitModel, PredictBatch, InplacePredict, Slice, serialization) into GBTree; stripped Dart down to a compatibility wrapper.
tests/cpp/gbm/test_gbtree.cc	Added GBTree_DropoutJsonIO test verifying gbtree with rate_drop saves/loads correctly.
doc/parameter.rst	Renamed DART section to "Additional dropout parameters for tree boosters" and removed DART-specific predict note.
doc/tutorials/dart.rst	Updated tutorial to describe dropout as a mode of gbtree rather than a separate booster.

---

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