Evaluate performance on few shot for an unseen tool after training toolformer

[dmahan93]

Should compare with a few items:

Generation with few-shot on base model, same as data generation
Generating with few-shot on base model, with similar setup to inference
Generation with few-shot on trained model

Not sure if there should be an additional comparison with data generation hyperparams (pull out API calls > 10%, evaluate k calls, M shot) for the trained toolformer, as well.