Catboost

Bugs fixed:
- `System of linear equations is not positive definite` when training MultiClass on Windows, 1022

Improvements:
- Massive 2x speedup for `MultiClass` with many classes
- Updated MVS implementation. See _Minimal Variance Sampling in Stochastic Gradient Boosting_ by Bulat Ibragimov and Gleb Gusev at [NeurIPS 2019](https://neurips.cc/Conferences/2019)
- Added `sum_models` in R-package, 1007

Bugs fixed:
- Multi model initialization in python, 995
- Mishandling of 255 borders in training on GPU, 1010

Improvements:
- New visualization for parameter tuning. Use `plot=True` parameter in `grid_search` and `randomized_search` methods to show plots in jupyter notebook
- Switched to jemalloc allocator instead of LFalloc in CLI and model interfaces to fix some problems on Windows 7 machines, 881
- Calculation of binary class AUC is faster up to 1.3x
- Added [tutorial](https://github.com/catboost/tutorials/blob/master/convert_onnx_model/tutorial_convert_onnx_models.ipynb) on using fast CatBoost applier with LightGBM models

Bugs fixed:
- Shap values for `MultiClass` objective don't give constant 0 value for the last class in case of GPU training.
Shap values for `MultiClass` objective are now calculated in the following way. First, predictions are normalized so that the average of all predictions is zero in each tree. The normalized predictions produce the same probabilities as the non-normalized ones. Then the shap values are calculated for every class separately. Note that since the shap values are calculated on the normalized predictions, their sum for every class is equal to the normalized prediction
- Fixed bug in rangking tutorial, 955
- Allow string value for `per_float_feature_quantization` parameter, 996

Improvements:
- For metric MAE on CPU default value of `leaf-estimation-method` is now `Exact`
- Speed up `LossFunctionChange` feature strength computation

Bugs fixed:
- Broken label converter in grid search for multiclassification, 993
- Incorrect prediction with monotonic constraint, 994
- Invalid value of `eval_metric` in output of `get_all_params()`, 940
- Train AUC is not computed because hint `skip_train~false` is ignored, 970

Bugs fixed:
- Incorrect estimation of total RAM size on Windows and Mac OS, 989
- Failure when dataset is a `numpy.ndarray` with `order='F'`
- Disable `boost_from_average` when baseline is specified

Improvements:
- Polymorphic raw features storage (2x---25x faster data preparation for numeric features in non-float32 columns as either `pandas.DataFrame` or `numpy.ndarray` with `order='F'`).
- Support AUC metric for `CrossEntropy` loss on CPU
- Added `datasets.rotten_tomatoes()`, a textual dataset
- Usability of `monotone_constraints`, 950

Speedups:
- Optimized computation of `CrossEntropy` metric on CPUs with SSE3

New features:
- Sparse data support
- We've implemented and set to default `boost_from_average` in RMSE mode. It gives a boost in quality especially for a small number of iterations.

Improvements:
- Quantile regression on CPU
- default parameters for Poisson regression

Speedups:
- A number of speedups for training on CPU
- Huge speedups for loading datasets with categorical features represented as `pandas.Categorical`.
Hint: use `pandas.Categorical` instead of object to speed up loading up to 200x.