EAGLE 3.1

Faster & resillient speculative decoding for LLM inference acceleration. Released fastest speculators for Kimi K2.6 and GPT-oss 20B/120B with _TokenSpeed_ team.