fix: upgrade deprecated LightGBM API in docs and version constraints

[Bortlesboat]

Summary

• Update documentation (docs/start/integration.rst) to use LightGBM's modern callback-based API instead of deprecated parameters
• Set minimum lightgbm>=3.3.0 in pyproject.toml (callback API was introduced in 3.3.0)
• Update example requirements.txt files from lightgbm==3.1.0 to lightgbm>=3.3.0

Problem

The core model code was already fixed in PR #974 to use callbacks (lgb.early_stopping(), lgb.log_evaluation(), lgb.record_evaluation()), but:

1. The documentation still showed the deprecated API (early_stopping_rounds, verbose_eval, evals_result passed directly to lgb.train()), which was removed in LightGBM 4.0
2. The minimum version in pyproject.toml had no lower bound, allowing installation of LightGBM versions that don't support the callback API
3. Example requirements.txt files pinned lightgbm==3.1.0, which predates the callback API (introduced in 3.3.0)

Changes

File	Change
docs/start/integration.rst	Replace deprecated lgb.train() parameters with callback equivalents
pyproject.toml	"lightgbm" → "lightgbm>=3.3.0"
examples/benchmarks/DoubleEnsemble/requirements.txt	lightgbm==3.1.0 → lightgbm>=3.3.0
examples/hyperparameter/LightGBM/requirements.txt	lightgbm==3.1.0 → lightgbm>=3.3.0

Test plan

• Verified core model files (gbdt.py, highfreq_gdbt_model.py, double_ensemble.py) already use callback API correctly
• Documentation examples now match the actual implementation in qlib/contrib/model/gbdt.py
• Version constraint ensures compatibility with the callback-based code

Closes #904

🤖 Generated with Claude Code

[Bortlesboat]

@microsoft-github-policy-service agree