Before this change the PEFT-wrapped policy did not save the policy's config
alongside the adapter config / weights which prevented us from changing the
policy config. Now the policy config is saved both in full training and PEFT
training.
This change makes loading the PEFT policy adapter much easier as well.