Model returned by inform stage only uses fraction of training data in final model

[sschmidt23]

Irene asked a question about the splitting of data that is done in Inform_FZBoost to split data to create a separate validation sample used in determining the best bump_thresh and sharpen params. Right now I have things set to train on a fraction trainfrac of the data, and use (1-trainfrac) fraction to compute cde loss values for a grid of bump_thresh and sharpen values. The model returned is the one trained on only trainfrac fraction of the training data. Re-computing the model with the full dataset would result in a better model, but would almost double the runtime of the inform stage. Maybe a compromise would be adding a config option named something like rerun_full that gives the user the option of recomputing the model on the full dataset.