major refactoring: process and split data properly, save splitted data

[MinaAlmasi]

Summary

1. Cleaned AI datasets prior to combining with human datasets to streamline pipeline, dropping any generations below min length prior to extracting metrics (closing issue Refactoring: Weird steps in the pipeline (cleaning at various steps that could be streamlined) #61)
2. Re-extracted metrics on cleaned AI datasets, computed perplexity with GPT-2 on both human and AI (took 5-8 hours to run GPT-2)
3. Deleted scripts insrc/utils e.g., process_generations and process_metrics that were used EVERYWHERE and instead did this processing ONCE (see point 4)
4. Introduced functionality of src/utils/process_X scripts to scripts in src/make_dataset where both text and metrics datasets are created in identical train, val, test splits and saved to datasets_complete (closing issue Create script to split data for classifiers and save to folder #70)
5. Added check_datasets to check that make_dataset/text_dataset.py and make_dataset/metrics_dataset.py make identical splits in terms of ID/MODEL (important for comparison between XGBOOST and TFIDF)
   6.Renamed datasets to datasets_files to complement newly added datasets_complete folder to handle the distinction between the individual files (generated text and extracted metrics) and the combined, splitted data (train/val/test text and metrics)
6. Updated PCA and classifier scripts (src/pca and src/classify) to handle new datasets
7. Plotted loadings for ALL 64 PC components
8. Created new heatmaps (metrics/heatmaps.py for ECHO presentation)

Future

New feats:

1. Reverse-engineer feature importances
2. Shap values

Refactoring:

1. Consider also saving PCA splits (right now they are transformed prior to classification)
2. Move more files from src/utils to their respective folder if not used in multiple places