Day-one model support — pre-warmed adapter registry for top-7 families

[caiovicentino]

When a new open-weight model drops, time-to-first-experiment should be measured in minutes, not days.

Scope: maintain openinterp_mcp.models.SUPPORTED registry with:

• Qwen3.5/3.6 (✓ already)
• Gemma-2/3/4
• Llama-3/4
• DeepSeek-R/V
• Mistral / Mixtral
• Phi-4
• Yi family

Each entry: layer mapping, d_model, tokenizer quirks, known activation positions, recommended SAE if any. Adapter PRs welcome.