Story for model-agnostic, predict-only "free" parameters

[justinxzhao]

Model-agnostic, predict-only parameters like top_k and threshold are "free" parameters since they have no bearing on how the model is architected or how the model is trained. However, they are interesting to configure and experiment with since they can have an outsized impact on evaluation metrics.

These parameters are currently set in the model config, but since they don't directly impact the model, it could be worth scoping out a more lightweight way to change these parameters without having to change the Ludwig config.