Mljar-supervised

Enhancements
- 179 add `need_retrain()` method to detect performance decrease
- 226 extract rules from decision tree
- 310 add support for MAPE
- 312 optimize prediction time
- 313 set stacking time threshold depending on best model train time
- 320 search for model with prediction time constraint
- 322 `n_jobs` as a parameter
- 328 disable stacking for small (nrows < 500) datasets

Bug fixes
- 214 move directory after training
- 246 raise exception when small time limit and no models are trained
- 247 proper display for optimize AUC and R2
- 306 add `mix_encoding` argument in `AutoML` constructor
- 308 fix dependencies error in kaggle notebook
- 314 bug fix in hill climbing in Perform mode
- 323 fix catboost bug with tree limit
- 324 325 bug for feature importance for small data

Many small improvements.

A lot of small tweaks and improvements :)

Enhancements
- 300 Add step with k-means additional features
- 299 Add Boost On Errors step
- 154 Sample weight available
- 229 Sort leaderboard (disabled for now for debug purposes)

Bug fixes
- 301 Fix storing unique keys in mljar tuner only for trained models
- 275 248 small fixes

Bug fixes
- 293 Typo in `is_scale_needed`
- 277 Fix problem with unit data
- 285 Restricted characters in LightGBM