server: in SSE mode, send HTTP headers when slot starts

[ngxson]

Overview

In stream mode, send HTTP header when the slot starts processing.

This may fix some timeout problems on clients. Ref: earendil-works/pi#5089 (comment)

Test script: https://gist.github.com/ngxson/d46506aab7e4a6b0f0df9f843d721cd6

On master, the header is sent after prompt processing completed (in this case: 0.8s from request sent):

[0.800s] STATUS: 200 OK
[0.800s] HEADERS:
[0.800s]   Content-Type: text/event-stream
[0.800s]   Keep-Alive: timeout=5, max=100
[0.800s]   Server: llama.cpp
[0.800s]   Access-Control-Allow-Origin: 
[0.800s]   Transfer-Encoding: chunked
[0.800s] --- streaming chunks ---
[0.800s] CHUNK: {"choices":[{"finish_reason":null,"index":0,"delta":{"role":"assistant","content":null}}],"created":1780083321,"id":"chatcmpl-eeaLzaFFF83iAY9uKZICpcMQsH8N5F3l","model":"unsloth/gemma-4-E4B-it-GGUF:Q4_K_M","system_fingerprint":"b9417-b5f52280f","object":"chat.completion.chunk"}

On PR, header is sent as soon as slot starts processing the prompt (0.008s from request sent):

[0.008s] STATUS: 200 OK
[0.008s] HEADERS:
[0.008s]   Content-Type: text/event-stream
[0.008s]   Keep-Alive: timeout=5, max=100
[0.008s]   Server: llama.cpp
[0.008s]   Access-Control-Allow-Origin: 
[0.008s]   Transfer-Encoding: chunked
[0.008s] --- streaming chunks ---
[0.799s] CHUNK: {"choices":[{"finish_reason":null,"index":0,"delta":{"role":"assistant","content":null}}],"created":1780083540,"id":"chatcmpl-oJ96sIgDZqP3a4YiY0wwx01FvCTJkGy6","model":"unsloth/gemma-4-E4B-it-GGUF:Q4_K_M","system_fingerprint":"b9419-b9d69e18e","object":"chat.completion.chunk"}

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: no

[ngxson]

hmm hold on, one test case failed

[ngxson]

@ggml-org/maintainers can I have the 2nd approval please 🙏