Remove unneed allocation

[roj234]

It seems that self.scores only changed when self._logits_all == True, so only allocate it when self._logits_all == True

[JamePeng]

You raised a good point. I've noticed that the old logic for loading and reading state in the original llama-cpp-python has deviated from the way llama.cpp operates. I'll probably add this to my to-do list to refactor this part of the logic.

You could also add a null check on the score array before it's called.

[roj234]

Unfortunately, (since I'm only used it to create embedding), I have not noticed that self.scores was used in generate (so it really need more refactors...)