llama.cpp's new CPU matmul speeds prompt processing 3-7x

A new llama.cpp CPU matmul path speeds prompt processing 3 to 7 times on x86 chips, merged into the project September 26. Prefill, feeding a model its context before it generates anything, ate most of the cost of CPU inference. This patch cuts that cost for long-context RAG pipelines and batch summarization jobs running on CPU-only boxes.

PR #27851 unpacks k-quant weights into 256x256 int8 tiles. A 16x16 VNNI microkernel runs over those tiles, replacing a vec_dot path that unpacked the same quantized weights repeatedly.

On an AMD 9950X3D running 8 threads at 8192x8192, q3_K matmul rose from 0.73 to 5.16 TFLOPS, about 7 times faster. q4_K reached 4.97 TFLOPS, up from 1.06, close to a 4.7x gain. Both roughly double the throughput of the existing repack path, and both show lower RMSE, so the speedup doesn't cost accuracy.

Token generation doesn't benefit: the new path engages only above 64 rows, and pure GEMV below that runs at 80% of stock speed.