Add support for fine-tuned models to be directly usable for prompt-tuning

[gkumbhat]

Description

Currently we only support loading of base models coming out from HF Hub or in native transformers format. But the models we generate from fine-tuning or after bootstrapping of TextGeneration module will be in "caikit" format, which will have slightly different folder structure.

This story is to add support to load these models that are in caikit format in the prompt tuning module's train and load functions.

[chakrn]

@gkumbhat is this done? Does the prompt tuning module look at finetuned models any differently than untuned models?

[chakrn]

@gkumbhat tagging to add some more details