model dropdowns list non-chat ids (embeddings/rerank/multimodal) that fail at runtime

Follow-up from PR #6578 (QwenCloud provider).

OpenAICompatibleProvider.list_models_for_api lists every id from GET /models unfiltered. DashScope's compatible-mode catalog is unusually mixed — a probe during review saw text-embedding-v3/v4, gte-rerank, qwen3-rerank, qwen-vl-max, qwen3-omni all delivered to the chat-model dropdown (13 of 51 ids non-chat). Selecting one fails at research time (now at least with the friendly model-not-found rewrite naming provider + endpoint).

Affects all OpenAI-compatible providers; most visible for QwenCloud.

Suggested: filter or annotate non-chat families in list_models_for_api (or client-side in formatModelsFromAPI), e.g. skip embedding/rerank/-vl-/-audio-/omni id shapes or tag them disabled. Also worth deduping and sorting catalog ids (duplicates currently produce duplicate dropdown entries and duplicate cache rows).