feat: curate HF model picker with OpenRouter analogues

[teknium1]

Summary

Follow-up to #3419. Instead of dumping 124+ models from the HF live API, the hermes model picker now shows 28 curated models organized by use case, with an Enter custom model name option for everything else.

Model categories

Category	Models	OpenRouter equivalent
Flagship	Qwen3.5-397B, DeepSeek-V3.2, Llama-3.3-70B, Kimi-K2.5, MiniMax-M2.5, GLM-5	qwen/qwen3.5, deepseek-chat, llama-3.3-70b
Reasoning	DeepSeek-R1, Qwen3-235B-Thinking, Qwen3-Next-Thinking, Kimi-K2-Thinking	deepseek-r1, qwen3-thinking
Coding	Qwen3-Coder-480B, Qwen3-Coder-Next, Qwen2.5-Coder-32B	qwen3-coder
Mid-size	Qwen3.5-122B, Qwen3-32B, Llama-3.1-70B, GLM-4.7, R1-Distill-70B	qwen3-32b, llama-3.1-70b
Small/Fast	Qwen3.5-9B, Llama-3.1-8B, Gemma-3-27B, Qwen3-8B, GLM-4.7-Flash, MiMo	llama-3.1-8b, gemma-3
Llama 4	Scout-17B, Maverick-17B	llama-4-scout, llama-4-maverick
Vision	Qwen3-VL-235B, Qwen2.5-VL-72B	qwen3-vl

How it works

Providers with 15+ curated models skip the live /models probe entirely. This keeps the picker clean and fast. Providers with small curated lists (z.ai, kimi, etc.) still query the live API as before.