Broker: stream responses through instead of buffering the full body

[adamcohenhillel]

The model broker buffers the entire upstream response before returning: response.read() at services/model-broker/openphone_model_broker.py:311. For stream:true Responses requests (the assistant's realtime path), this means the connection stalls for up to 120s (_proxy_to_openai timeout) and then dumps the buffered body — the assistant cannot stream tokens through broker mode.

Fix

• Detect Accept: text/event-stream or stream:true in the payload.
• Pass through the upstream response body as SSE chunks (urllib.response.HTTPResponse.read1 in a loop, flushing per chunk).
• Preserve Content-Type: text/event-stream, Transfer-Encoding: chunked, and disable buffering in the response wfile.
• Add a smoke test using a synthetic SSE upstream.

Size: M