Free up GPU memory after unregistering model in meta reference inference

[SLR722]

🚀 Describe the new functionality needed

as title

💡 Why is this needed? What if we don't build it?

This can free up the memory if we want to switch the models for inference. Also semantically, it doesn't make sense to continue occupying the memory after the model is unregistered

A real application is that if we want to inference the base model -> do post training -> inference the finetuned model. If we can not free up the memory occupied by the base model with unregister command, it's easy to be OOMed.

Other thoughts

No response