Is your feature request related to a problem? Please describe.
When using Llama.save_state, the size on disk is very large, partially because Llama.scores is always of size (n_ctx, n_vocab), even when the number of tokens actually used is much less than this.

Describe the solution you'd like

• Only store the number of tokens used by the model (self.n_tokens).
• If possible, re-use logits from the llama_context struct instead of serializing scores.
  • These didn't match when I loaded both and compared, though didn't have logit processor callbacks.

Describe alternatives you've considered

• Somehow compressing LlamaState
• Re-using logits from C++ state saving functions and eliminating scores on Python model
  • Looks like this is needed for sampling logic

Partially fixed by #1296