fix(explainability): support dynamic SHAP log-loss baselines

[jiangxt2]

Description

This PR fixes SHAP TreeExplainer log_loss explanations with SHAP 0.52.0 and simplifies the Tributo pull request template.

The SHAP fix handles the dynamic per-sample base_values returned by SHAP as callable objects, materializes them with the supplied labels, and fails closed for missing labels, mixed callable values, or unsupported shapes. This prevents the previous float + method failure during output reconstruction.

The PR template now follows Ray's concise Description, Related issues, and Additional information structure. Fixed test and open-source checklists were removed; contributors are prompted to record relevant test results, pending integration validation, user/API changes, dependency review, and known risks when applicable.

Related issues

None.

Additional information

• User-facing impact: SHAP tree explanations using output_target="log_loss" work with the current locked SHAP version; the template change only affects pull request authoring.
• Validation:
  • The original real XGBoost regression test passed: 1 passed.
  • tests/explainability/test_shap.py: 9 passed.
  • Ruff check/format, CI test-plan audit, and git diff --check passed.
  • Commit hooks passed for the changed Python files and the template change.
• Environment limitation: the full scheduled data/training shard and the precheck changed-test stage encountered an intermittent native macOS ARM XGBoost process segmentation fault (exit 139); this was not a pytest assertion failure. The targeted real regression test passed independently.