llama-cli: fix not copying preserved tokens

[michaelw9999]

Overview

Llama-server will properly read preserved token IDs and copy chat_params.preserved_tokens into task.params.sampling.preserved_tokens. However, it appears this was left out of llama-cli which does not.
This PR fixes that.

I had just finished a new PR (about to post ) for the new Cohere2 MoE arch and BLS-Mini-Code-1.0 model, and kept fighting chat template issues repeatedly until I saw it only was happening in llama-cli testing, but was not a problem happening when trying via llama-server.
So some special tags, for example, <|END_THINKING|> were getting detokenized, so the model was never showing output when a response would start, leading to the response not showing. This only was in llama-cli.

Previously:

> Hi BLS

[Start thinking]
The user says "Hi BLS". Likely they are greeting. As an AI, we should respond politely. No policy issues. We can ask how can we help.Hello! How can I assist you with today?

was really because of: ... how can we help.<|END_THINKING|><|START_TEXT|>Hello! ...

After this fix:

> Hi BLS

[Start thinking]
The user says "Hi BLS". Likely they are greeting. As an AI, we should respond politely. There's no request for content. We can respond with a
[End thinking]

Hello! I'm Command, a Cohere-built language model. How can I assist you with your query today?

[ Prompt: 23.6 t/s | Generation: 195.3 t/s ]

Additional information`

Tested the fix with a few other models previously working and did not see any change in behavior.
test-chat, test-chat-auto-parser, test-chat-template, test-jinja, test-reasoning-budget, test-sampling all pass.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, used to find the real root cause of the issue and propose a fix.

[CISC]

@michaelw9999 Thanks for the reminder.