Add supervised methods

[ArtemVazh]

1. Added support for supervised methods: SAPLMA, Sheeps, MIND, SATRMD, LookBackLens, and TAD. To adjust the target metric used during training, you can modify the configuration by setting stat_calculators.5.cfg.target_metric=AlignScore.
2. Introduced separate calculators for attention and embedding processing: These calculators receive raw data from greedy_probs.py and process it independently.
3. Added layer control for supervised and density-based methods: Added a layer (or layers) parameter to specify the required layers for each method. This enables layer selection in the manager prior to execution, allowing extraction data only from necessary layers. This is required since current calculators still rely on static meta_info and do not support dynamic layer-specific statistics.
4. Improved statistic sorting logic: Training statistics are now initialized only after inference statistics. This enables specifying parameters in the builder environment that determine which statistics are initialized and needed for the training run - preventing unnecessary data extraction.
5. Minor fix: Disabled verbose output in AlignScore.
6. To use supervised methods, set the parameter estimators=default_estimators_with_supervised
7. To set the threshold for the quality metric in the supervised method, set the parameters estimators.42.cfg.metric_thr=0.85 estimators.43.cfg.metric_thr=0.85 estimators.44.cfg.metric_thr=0.85 estimators.45.cfg.metric_thr=0.85

[ArtemVazh]

estimators.49.cfg.metric_thr=0.85 estimators.50.cfg.metric_thr=0.85 estimators.51.cfg.metric_thr=0.85 estimators.52.cfg.metric_thr=0.85

[IINemo]

Could you please update the pull request, for compatibility with main?