Support of ML runtime using KML-format

[mkornaukhov03]

Common

Add new subsystem into runtime, it is inference of machine learning models. We have developed new format -- KML, this PR add a support of models in such a format. For now, it only supports xgboost and catboost models.

API

PHP side API is following:

function kml_xgb_predict(string $model_name, array $features): array< float >;
function kml_catboost_predict(string $model_name, array $float_feats, array $cat_feats) : array< float >;
function kml_catboost_predict_multi(string $model_name, array $float_feats, array $cat_feats) : array< array< float > >;

To load KML-models (.kml files), you should specify kml-dir option.

Internals

Master-process iterates specified directory recursively and load all .kml files inside it simultaneously calculating the least sufficient amount of memory that is used to predict a model. The calculated value is used to allocate buffer of value size (one buffer per worker). This buffer is going to be used as arena for each call of kml_xgboost_predict, kml_catboost_predict, kml_catboost_predict_multi, that is why there are no heap allocation during script working time.