[ENH] Finetuning hyperparameter changes in Global API [draft]

[felipeangelimvieira]

This is a draft to register and collect ideas for supporting finetuning in the new Global API

Problem

As global forecasters are moved to the new API, we also need to support updating hyperparameters for finetuning purposes. However, using set_params would reset the model and erase the pretained state.

I can think of two use cases that should be taken into consideration:

1. User wants to load a global model and change its hyperparameters for a finetuning step
2. User wants to set new parameters of a global model and reset it (e.g. by change the model name / version)

And possible edge cases:

1. User tries to change hyperparameters that would not make finetuning possible (e.g. changing the base model version)
2. User tries to change finetuning parameters of a global model inside a meta-model (composite estimator or pipelines). They might want to only set finetuning hyperparameters without resetting the global model, but they might also try to set parameters of adjacent models that do not support pretraining.