llama.cpp b10208 doubles Intel GPU prompt processing with oneMKL flash attention

ggml-org has shipped llama.cpp b10208, a significant SYCL backend update that brings oneMKL GEMM flash attention to Intel GPUs. The new kernel routes multi-token prefill through XMX-accelerated matrix operations, replacing the slower TILE path. In testing on B70/Battlemage hardware with 32K context and q8_0 KV cache, Gemma-4-26B jumped from 746 t/s to 1473 t/s (1.97x), while Qwen3.6-27B improved from 330 t/s to 609 t/s (1.85x). The release also expands support to all KV cache types — F16, BF16, F32, and quantized — since the kernel converts non-F16 data to F16 before GEMM, eliminating a previous restriction.

The update fixes a critical flash-attention layout bug where the kernel wrote head outputs using a dense head-major layout instead of the interleaved layout expected by TILE's combine function. This corrupted attention for all models except Qwen3.6-27B, whose sparse GQA heads avoided visible overlap. The code also removes redundant SYCL stream waits, adds a GGML_SYCL_ENABLE_MKL_FA environment variable for A/B testing, and gates the MKL path behind a Q->ne[1] >= 32 threshold to keep token generation on the fused VEC kernel. Debug and diagnostic flags (GGML_SYCL_MKL_FA_DEBUG, GGML_SYCL_MKL_FA_DIAG) are included for profiling.