README: note input embedding is loaded from verifier (not shipped), checkpoint ~3.7GB

Remove redundant embed_tokens (shared from verifier); 6.03GB->3.68GB. vLLM deepseek_eagle3 loads it via includes_embed_tokens=False path.

Add Kimi-K2.7-Code Eagle3-MLA draft (full vocab) + README with per-benchmark accept_length

initial commit