model picker: filter out non-chat models (embeddings, speech, image, rerankers)

The /model picker was flooded with hundreds of non-chat models because
OpenAI-compatible profiles (NVIDIA NIM, FPT, Chutes, Groq) and Bedrock
expose their entire catalogs (embeddings, whisper/ASR, TTS, image
upscalers, rerankers, OCR/parse models, etc.).

Add model_name_is_likely_non_chat() heuristic and apply it via
is_listable_model_name plus a final retain pass in
multiprovider_model_routes so the snapshot/catalog used by both
available_models_display and the model routes is filtered at the source.

Conservative, token-boundary aware matching keeps multimodal chat models
(e.g. llama-3.2-90b-vision-instruct, Qwen2.5-VL-Instruct).