LightGBM

Feat/optimize single prediction (#2992)

* [performance] Add Fast methods to C API for SingleRow Predictions

 * Add methods to C API to make single-row predictions faster:

   - LGBM_BoosterPredictForMatSingleRowFastInit (setup)
   - LGBM_BoosterPredictForMatSingleRowFast (predict)
   - LGBM_FastConfigFree (cleanup setup outputs)

* Code syle cleanup

* Fix lint errors

* [performance] Revert FastConfig improvement to pass data at init

This reduces optimization by 5% / 30% with this branch but makes it so it can be used for higher level wrappers in MMLSpark.
And outside it as well.

* [performance] Introduce Fast variants for SingleRow predictors.

Although this already provides performance gains by itself for any
callers, two new functions were added to Java's SWIG interfaces to
exploit that AND the GetPrimitiveArrayCritical data fetches.

* [tests/profiling] Profile Fast predict methods

Build with -DBUILD_PROFILING_TESTS=ON and copy the default
model trained on the Higgs dataset from the benchmarks repo

 https://github.com/guolinke/boosting_tree_benchmarks.git

to LightGBM repo root and run the lightgbm_profile_* binaries.

The single instance used is the first row from that dataset.

* Update comment on CMakeLists.

* Fix doxygen-introduced issue (#threads)

* Fix conflicts due to new RowFunctionFromCSR signature in master

* Change FastConfig ncol to int32_t.

* Removed profiling folder

* fix doxygen typo include/LightGBM/c_api.h

Co-authored-by: Nikita Titov <nekit94-08@mail.ru>

* fix doxygen typo include/LightGBM/c_api.h

Co-authored-by: Nikita Titov <nekit94-08@mail.ru>

* fix doxygen typo include/LightGBM/c_api.h

Co-authored-by: Nikita Titov <nekit94-08@mail.ru>

* Doxygen: change new docstrings to double back-quote

Co-authored-by: alberto.ferreira <alberto.ferreira@feedzai.com>
Co-authored-by: Nikita Titov <nekit94-08@mail.ru>