Cosmos3-edge PR 5: end a failed chat stream with an error event instead of stop

[merceod]

APIServer.iter_result_chunks already reports a request that fails after its stream opened (an engine error mid generation, a delivery timeout) as a terminal error chunk, and /generate and /v1/videos forward it. The /v1/chat/completions SSE path dropped it and closed with finish_reason: "stop", so a truncated answer read as a complete one. I ran into this while benchmarking concurrent Cosmos3-Edge reasoner requests, where worker errors showed up as short answers with a normal stop and the throughput numbers counted them.

The stream now ends with an OpenAI-style error event and [DONE], no stop chunk, matching the router's non-streaming error body.

data: {"error": {"message": "Error in worker: ...", "type": "server_error", "code": 500}}

data: [DONE]

Test in test/modular/test_openai_router.py::test_chat_stream_reports_a_failed_request_in_band (stubbed API). ruff and test/modular pass.