speculative decoding silently broken for Qwen3.6 on the TurboQuant fork — PR to fix

if you're running Qwen3.6-35B-A3B on the TurboQuant fork and you tried speculative decoding, it was quietly doing nothing. the server just falls back to normal decoding without any error. basic idea for anyone unfamiliar: you run a tiny model (like Qwen3.5-0.8B) alongside your big model. the small one guesses the next bunch of tokens really fast, then the big model checks all the guesses in one pass. whatever the big model agrees with, it keeps — whatever it rejects, it redoes. so if your big m