Update all w8a8ch definitions' correctness reference, aligns with what llama.cpp low-bit kernel correctness reference style: dequant input, then calculate in full precision

Definitions, solutions, and evaluation trace updates for ncnn kernels

add baseline solution source at all neural network kernel definitions

Update definition and workload for CV model kernels:

local snapshot before syncing with HF