only in new data flag

[Max Kuhn]

Sort of like skipping in recipes, can we put a flag in adjust_predictions_custom() to apply it to new data (i.e. not during the model optimization/evaluation cycle)?

I'd like people to avoid exponentiating predictions prior to metric computations. We could add a similar flag to fit.tailor() that says, "We are in pure new sample prediction mode." We can set that appropriately in tune.

[Simon P. Couch]

Related, maybe duplicate, of #18.