Expose disabling decoder loading for classification and regression tasks

[lukasugar]

Cls/reg fine-tuning previously could not override load_decoder even though BaseModule supported it. ClassificationModule and RegressionModule both go through ClsRegBase, and ClsRegBase.__init__ did not accept or forward a load_decoder argument. As a result, BaseModule always used its default load_decoder=True for cls/reg checkpoint loading.

This change makes the option explicit for cls/reg fine-tuning by adding load_decoder to ClsRegBase, forwarding it to BaseModule, adding training.load_decoder: True to the default cls/reg fine-tune configs, and passing that config value from finetune_cls.py and finetune_reg.py.

The default behavior is unchanged: cls/reg still load decoder weights by default. The difference is that users can now override the behavior from config or CLI, for example:

asp_finetune_cls training.load_decoder=False
asp_finetune_reg training.load_decoder=False

Imo, it makes sense to not load decoder in many classification and regression cases.

[Sllambias]

Hi Luka,
Thanks for your PR. I think you are right that it makes sense to expose this, so thank you for contributing with the changes, and since this is completely backwards compatible I say we get this merged right away

[lukasugar]

Great! I don't have merge permission, so please merge it.

I can only comment on the PR: