Update stale docstrings, docs and tutorial outputs (#77)

[upadhyan]

Closes #77.

Summary

• Docstrings (sgtlearn/)
  • Documents tao_n_runs and tao_lambda on all four estimators.
  • States that feature importances raise after the default TAO fit.
  • Describes num_partitions as a maximum.
  • Notes that criterion also grows the inner trees and accepts "log_loss", and that max_features samples logical features.
  • Describes current regression seeding and MAE coordinate-descent gating.
  • Corrects pair-node NaN routing.
  • Adds full TAO_refine Parameters/Returns/Raises, noting that the cost is lambda_ * n_node with early stopping.
  • Adds plot_tree parameter semantics, feature_names_in_, the tree_export keys, and the missing predict/predict_proba docstrings.
  • Removes references to things that don't exist (fit(column_names=...), the private _export.plot_tree path).
• Read the Docs pages
  • tao.rst: n_node cost, tao_pair_scale in the knob lists, candidate scope, early stopping.
  • estimators.rst / ensemble.rst: pairwise_penalty as the γ growth cost, λ disambiguated from tao_lambda, predict shapes.
  • Quickstart and installation (scikit-build-core, best-effort stubs).
  • Roadmap: v0.3.1 and Unreleased (export_text) sections.
  • Fixed about 20 broken cross-references.
  • conf.py: dropped the global inherited-members: False, which overrode :inherited-members: and hid fit/predict on the forest pages. Inherited sklearn methods now render too.
• READMEs, examples, C++ docs
  • Build and layout sections corrected.
  • Binding docstrings and the regression header comment now match the code (comments only, no logic changes).
• Notebooks
  • All 10 re-executed. Outputs are identical to v0.3.1; only the execution metadata changed, and split stdout blocks were merged.

Notes

• The tutorial prose was audited separately; that report is delivered outside this PR. Its findings (stale numbers in tao/sgt-k, etc.) are not addressed here.
• Code bugs found during the audit are tracked in TAO at lambda_=0 can lower training accuracy for classifiers with 3+ children per split #80 (TAO can lower accuracy at λ=0 for K≥3), feature_dict int key equal to an unlisted column index silently drops that group #81 (feature_dict int keys drop columns) and plot_tree crashes on SGTRegressor with max_depth truncation or multi-output y #82 (plot_tree crashes on truncated or multi-output regressors). tao.rst keeps the "accuracy does not decrease at λ=0" guarantee, which TAO at lambda_=0 can lower training accuracy for classifiers with 3+ children per split #80 restores.
• ruff format --check already fails on sgtlearn/_multioutput.py on main; this PR leaves it untouched.

Test plan

• sphinx-build -W --keep-going -b html docs passes. With -n, no attr/meth/obj/func/mod reference warnings remain.
• pytest tests: 1241 passed, 2 skipped.
• ruff check sgtlearn is clean.
• Extension rebuilt with the edited binding docstrings; re-executed notebooks give identical outputs.

🤖 Generated with Claude Code