llama.cpp PR makes Q2_0 3.0–3.6x faster on x86 CPUs

llama.cpp PR #26348 adds an x86 VNNI implementation for the Q2_0 × Q8_0 dot product. The author's controlled CPU-only benchmarks show 3.0–3.6x higher throughput, with 8B decode rising from 2.39 to 8.20 tok/s.