llama.cpp PR makes Q2_0 up to 3.6x faster on x86 CPUs

llama.cpp PR #26348 adds an x86 VNNI implementation for the Q2_0 × Q8_0 dot product, delivering 3.0–3.6x higher throughput in CPU-only benchmarks; an 8B model's decode speed jumps from 2.39 to 8.20 tok/s.