fix: /v1/embeddings now returns OpenAI shape (OQP-4)

[TadMSTR]

Summary

• dispatch_request treated any response with transfer-encoding: chunked as streaming, returning StreamingResponse instead of JSONResponse
• Ollama's /api/embed sends application/json with chunked TE — so the OpenAI compat layer's isinstance(response, JSONResponse) guard in proxy_handler always failed
• wrap_response was never called; callers got the raw Ollama-native shape {embeddings, model} instead of the OpenAI shape {object, data, model, usage}
• Fix: remove transfer-encoding: chunked from the is_streaming heuristic — chunked TE is a transport mechanism, not an application-level streaming indicator. Streaming is identified solely by content-type (text/event-stream, application/x-ndjson)

Test plan

• test_chunked_json_response_not_treated_as_streaming — new regression test
• All 139 existing tests pass
• Live verify after container rebuild: /v1/embeddings returns OpenAI-shaped response

Closes OQP-4

Generated with Claude Code