llama.cpp PR: up to 169% faster quantized-KV decode on Intel Battlemage

PR #26689 reports up to 169% faster decode at 118K context with q4_0/q8_0 quantized KV caches. The gain comes from switching a SYCL FlashAttention dispatch from the VEC kernel to TILE on Intel Battlemage.