#2154 [Feature]: answer GET /models/sse (llama.cpp's model status feed)

Descrição

Strata already mimics llama-server's router on `/models`: the listing reports the model's status (`loaded` / `unloaded`), so the model loads on use and can unload between requests. A front-end that watches that status (llama.cpp's own web UI, and any client that follows llama-server's model management API) also subscribes to `GET /models/sse` to learn about load and unload events as they happen, instead of polling. Strata answers that with a 404 today, so those clients log a warning and reconnect forever - Zed's llama.cpp provider retries every 5 s - and their model list never refreshes on its own.

The ask: answer `GET /models/sse` with 200 and `text/event-stream`, keep the connection open, and emit llama.cpp's `status_change` events (`loading`, `loaded`, `unloaded`) when the model loads, unloads or restarts - all transitions the server already knows about. A quiet feed sends the usual SSE-comment keep-alive.

I have a change ready that does this (the route, a broadcast on the existing load/unload transitions, tests in `serve/test_lifecycle.py`, docs in `DETAILS.md`); the PR follows. It is purely additive: every existing endpoint answers exactly as before.