Llama.cpp triples Q8_0 speed on Intel Arc

This optimization is a massive win for the Intel Xe2 ecosystem, finally unlocking the bandwidth potential of Battlemage GPUs for high-precision local LLM inference.

• –Q8_0 quantization previously suffered from non-coalesced memory access due to its non-power-of-two 34-byte block size.
• –Reordering allows for efficient GPU cache usage, pushing performance on Qwen3.5-27B from 4.88 t/s to over 15 t/s.
• –The speedup makes Q8_0 quantization faster than the less-precise Q6_K on Intel hardware.
• –The fix also addresses a silent bug in SYCL buffer initialization that prevented the optimization from being enabled for Q8_0 tensors.
• –Verification included binary-patching Intel's closed-source IPEX-LLM to confirm the hardware's theoretical limits before implementing the open-source fix.