[Feat]: Simpler new model support

[dxqb]

Describe your use-case.

With the amount of new models released, it could be easier to implement new models:

major:

• code deduplication in modules/dataLoader: cache_modules, output_modules, etc.
• code deduplication in modules/modelLoader
• the samplers in modules/modelSampler duplicate code for each model in modules/BaseModelSetup (how to call the transformer, creating ids, etc.)
• deduplicate the debug code in modules/BaseModelSetup...
• ...and setup_optimizations
• find a better way to select modules in create.py
• move the LoRA key conversion code back into OneTrainer https://github.com/Open-Model-Initiative/OMI-Model-Standards/tree/main/src/omi_model_standards/convert/lora

major but difficult:

• the mgds modules that have to be created for tokenization and text encoder, duplicate code in modules/model.encode_text for each model

minor:

• boilerplate code in modules/modelSaver
• some boilerplate code in modules/model
• boilerplate code in modules/*LoRASetup and *FineTuneSetup

Not exactly sure how to go about this, because changing these things can break existing models and therefore requires lots of testing.

What would you like to see as a solution?

see above

Have you considered alternatives? List them here.

keep everything as-is, or only do the easy points above