Stream tokens in `nanotune chat` instead of waiting for the full reply

[will-lamerton]

Problem

nanotune chat waits for the entire response before rendering anything. chatCompletion() in src/lib/llama-cpp.ts does a plain fetch and awaits the full JSON body.

On a long reply the UI sits on a "Generating..." spinner for many seconds with no indication of progress. Small local models are exactly the case where you want to see tokens land — it's how you tell the model is working versus stuck.

llama-server already supports SSE streaming on /v1/chat/completions with "stream": true.

Proposal

• Add a streaming variant alongside chatCompletion (keep the non-streaming one — benchmark doesn't need streaming and relies on the final timings block)
• Render tokens into the chat transcript as they arrive
• Keep the existing per-turn stats line; SSE responses still carry timings in the final chunk

Acceptance criteria

• Tokens appear in nanotune chat as they're generated
• TTFT / tok/s / token count stats still shown after each turn
• Interrupting mid-generation cleanly cancels the in-flight request
• benchmark behaviour is unchanged