Arc Pro B60/B70 llama.cpp GGUF Quantization Optimization

Measured on Intel Arc Pro B70 (Battlemage, 608 GB/s theoretical bandwidth), Qwen3.5-27B, one GPU: Q4_0 23.67 t/s (57% of peak), Q4_K_S 23.05 (56%), Q4_K_M 20.56 (53%), IQ4_XS 17.52 (40%), Q6_K 13.83 (48%), Q5_K_M 13.78 (41%), IQ4_NL 5.85 (14%), Q8_0 4.88 (21%). The decisive datapoint: IQ4_NL and Q4_0 are the same size (14.60 vs 14.63 GiB) and differ 4x in speed, which excludes data volume and points at kernel efficiency. Root cause: the SYCL backend's reorder optimization (splits scale factors from weight data for coalesced memory access) was implemented for Q4_0, Q4_K and Q6_K in PR #12035 but never extended to Q8_0 or IQ4_NL. Fix for Q8_0 landed as PR #21527. On Arc Pro B60 (456 GB/s), the same pattern holds but at lower absolute throughput. The generalizable lesson: on a new Intel dGPU, assume some quant types have no tuned kernel until measured, and prefer Q4_K_M as the safe default. Context KV size: default context is too big and causes OOM; use -c 8192 or smaller. Batch size: larger batches improve throughput but require more VRAM for KV cache. The Vulkan backend shows the same quantization performance pattern as SYCL, confirming it is a kernel efficiency issue, not a backend issue.