Fix accelerator issue

[SachiraKuruppu]

• Provide the complete model (rag_model) to the accelerator instead of giving the retriever and the generator separately.
• Change the forward path to use the retriever or the generator based on the task string.

This declutters the training code.

[SachiraKuruppu]

This needs a bit more work:

• Add arguments to AutoModelForRagE2E to load the model for inference with peft layers.
• Fixdalm/eval/eval_with_hnsw.py use of rag_model.forward(...) function.

I'll have a look tomorrow, or feel free to fix these.

[shamanez]

looking good