Add measured M5 Pro decode (~62 tok/s, TinyLlama-1.1B fp16)

LLM text-generation on the ANE (usage card)

initial commit