Categorical data support (part 2)

[trivialfis]

This is a follow-up to #6503 .

Hyper-parameters

• Large scale benchmarks.
• Investigate different parameters for preventing overfitting. (Implement max_cat_threshold for CPU. #7957)
• Compare results between different implementations.
• Produce documents about how each parameter affects model performance.

Performance

• Optimize GPU evaluation function.
• Optimize CPU evaluation function by iterating only half of the histogram.
• Better handling for dataframe Avoid dataframe deep copy while creating dataset microsoft/LightGBM#4596 Support primitive types of pyarrow-backed pandas dataframe. #8653.

Features

• Investigate the potential support for the exact tree method.
• Possible roadmap for other algorithms.
• Missing value handling. (Incorrect hanlding of missing value for categorical features. #7917, Handle missing categorical value in CPU evaluator. #7948, Unify the cat split storage for CPU. #7937, Handle missing values in one hot splits. #7934, Add max_cat_threshold to GPU and handle missing cat values. #8212)
• Add demonstration to JSON model parser. (Update JSON parser demo with categorical feature. #8401)
• PySpark ([pyspark] Categorical data support. #8398)

3-party libraries

We should have a simple function that's similar to our JSON parser example to export a Python model in XGBoost.

• SHAP Xgboost binary format compatibility shap/shap#1302 SHAP TreeExplainer is not compatible with XGBoost models that use Pandas categorical features shap/shap#2662

Language bindings

• R. Need to support factor. ([R] Implement feature info for DMatrix. #8048, Support dataframe data format in native XGBoost. #9828)
• Java ([jvm-packages] add jni for setting feature name and type #7966)
• Scala & Spark

Suggestions/discussions are welcomed!