docs(readme): routing tier table matches what the router picks today

[VickyXAI]

Doc-only. I ran route_with_catalog against the live /v1/models (2026-10-02):

• The auto MEDIUM tier picks google/gemini-3.5-flash for code and google/gemini-2.5-flash for plain chat. It no longer picks moonshot/kimi-k2.7, which is now a hidden id. There is now a note that the task ranker can put a fallback ahead of the configured primary.
• For "What is 2+2?", the free example now lands on nvidia/llama-3.2-11b-vision, not nano-omni.
• The nvidia/deepseek-v4-pro note now names the right redirect target. The gateway sends it to a Nemotron model, not V4 Flash.
• The premium openai/gpt-5.3-codex example is correct (it is what the router picks today and it is live), so it stays.

Refs BlockRunAI/blockrun#781