Eliminate text parsing from feature importances and evaluation metrics

[hcho3]

Currently, important functions such as feature importances and evaluation metrics rely on parsing of text strings, specifically the text output from the model dump function. For example:

https://github.com/dmlc/xgboost/blob/68c55a37d9bb680fe435f1d011e5fea62be97d22/python-package/xgboost/core.py#L1797-L1832

https://github.com/dmlc/xgboost/blob/68c55a37d9bb680fe435f1d011e5fea62be97d22/jvm-packages/xgboost4j/src/main/java/ml/dmlc/xgboost4j/java/Booster.java#L509-L540

https://github.com/dmlc/xgboost/blob/68c55a37d9bb680fe435f1d011e5fea62be97d22/python-package/xgboost/training.py#L85-L91

https://github.com/dmlc/xgboost/blob/68c55a37d9bb680fe435f1d011e5fea62be97d22/jvm-packages/xgboost4j/src/main/java/ml/dmlc/xgboost4j/java/Booster.java#L240-L255

Also see https://github.com/dmlc/xgboost/issues/4665#issuecomment-532932603 https://github.com/dmlc/xgboost/issues/4665#issuecomment-532945623

We should aim to eliminate all such uses of text parsing, since a slight change in the text dump will cause all these functions to break.

Proposed replacement:

• Feature importances: Implement a new C++ function that returns a JSON string representing features and their importances.
• Evaluation metrics: Implement a new C++ function that returns a JSON string representing eval set names and their eval metrics.

Now that we have a functioning JSON library as well as numeric printing function (charconv) in XGBoost, it should be doable.

[trivialfis]

Looked into this a little bit. The implementation isn't difficult, but depends on https://github.com/dmlc/xgboost/pull/6605 due to the use of feature names/types. I will try to figure out a better way to store those information.