Performance
* perf: minor performance improvements for arg extraction and frame detection (15)
* save best model chkpt based on val loss
* Adding loader setup to training
* add more optional config for loggers/callbacks during training
* adding explicit logging for test/train/val loss to end of epochs
* rever to default PL logging behavior if no loggers are provided
* adding helpers for model evaluations
* try to standardize arg extraction output
* standardize punct in args extraction
* use fast tokenizer for sent clenanup
* switch to just using tokenizer cleanup for speed
* run clean_up_tokenization just once before arg extraction, not for each arg
* fixing val_metrics err ([`7e03969`](https://github.com/chanind/frame-semantic-transformer/commit/7e039695b8cfa1ae5b8de1b9b4d4145e5bec3884))