Optimize llama.cpp? Hmm.

Wrt speculative decode, basically zero finetunes keep it. I'm testing with some ridiculous abliterated amalgamation so spec decode has been gone for most of its ancestry.