perf: maddubs kernel + nrc=4 multi-row for Q1_0_g128 (3.5-3.75 t/s)

perf: optimized AVX2 kernel + COM6-inspired matmul dispatch (0.2 -> 3.43 t/s)

Q1_0_g128 CPU kernel fix + AVX2 SIMD (fork of PrismML-Eng/llama.cpp)