speculative decoding in llama.cpp

[bytesupply76]

I think the bottleneck is somewhere else. I get why that seems right, but on A6000 the bottleneck felt elsewhere. Worth testing the simpler explanation before adopting the more interesting one.