[REFRACTOR] Cleanup backend and model_type usage

[LRL-ModelCloud]

No description provided.

[Qubitium]

@LRL-ModelCloud @ZX-ModelCloud I can't merge this yet. This introduces new backend property to BaseModel but the backend prop is not correctly updated to reflect the actual value during runtime.

For example, I load from_quantized model and set backend=auto, but since it is auto (not real), it will be converted to real backend after "auto" checking for compatible backend for this model. Once the backend is selected, the .backend prop value should reflect the real backend and not the incorrect auto.