Explaination of --model_prediction_type

Could you point out where's these setting from?
"raw" is easy to understand. For "additive" and "sigma_scaled", how is they introduced? Is them proposed just by this repo? What are they meant to benefit ?

Start by locating the --model_prediction_type setting in the Python sources and trace where raw, additive, and sigma_scaled are defined or used. Check any related documentation or history to identify their origins, intended benefits, and differences. Done means documenting those explanations clearly for users.

Written by the indexing model from the issue text.