create model_selection_score cmd line parameter with a few common available options

[tmills]

Potential enhancements:

• additional parameter to subset labels to average over (e.g., ignore majority label with very high score)
• allow strings that can be passed to HF evaluate functions

[tmills]

Proposed solution: Two new training arguments: --model_selection_score and --model_selection_label. Example: 'acc' or 'f1' as model selection score. In compute_metrics_fn, check these arguments and set one_score based on values of those 2 arguments. No argument: Macro f1 (current default), 'f1' with no label provided, macro f1, 'acc' with no label provided, accuracy, 'acc' with label provided, error?, 'f1' with label provided, look up label index in dataset and use that label's f1.

[tmills]

For cases where you may want average f1 across multiple labels, --model_selection_label could take a list.