fix(serve): accept chat completions without a model field

[tmark00]

The bundled llama.cpp webui never sends model on /v1/chat/completions — it treats the server as single-model and takes the display name from /props. The parser rejects that with

400 {"error":{"message":"missing required field: model","param":"model","type":"invalid_request_error"}}

so every message typed into the UI served by --webui fails. It is not a stale-cache or webui-version problem: I reproduced it on both published bucket versions (1789559630374 and 1789652224694), in two browsers, and with --model-id set to the id the UI had listed from /v1/models.

This PR makes an absent model leave the field empty; handle_chat_completions fills in public_model_id_ before validate_openai_model. Unchanged: a model that is present but does not match is still rejected, "model": "" and a non-string model still fail, and the Anthropic and Responses paths are untouched.

Verified on Windows 11 x64, RTX 5090 Laptop, against the QUASAR v3 artifact:

• webui at /: message sent, answer rendered, POST /v1/chat/completions returns 200 instead of 400;
• curl without the field returns "model":"qwen3.8-27b" in the response body;
• curl with a wrong model still returns model_not_found.

Builds on cometkim/dev @ 024f6a94. Independent of #2 (the Windows --webui download fix), though on Windows you need both before the UI can be reached at all.

🤖 Generated with Claude Code