llama.cpp adds SYCL Flash Attention and XIELU support for Intel GPUs

Recent updates to llama.cpp introduce several improvements for SYCL/Intel hardware, including performance boosts and new feature support.

• Flash Attention with the XMX engine via oneDNN graph API is now available for Xe2 architectures, significantly accelerating prefill speeds for models like Qwen3.6-27b-Q8_0.
• The codebase now supports the XIELU operation type and adds fp16 support for the conv2d_dw kernel.
• Fixes include increasing the minimum buffer size for USM system allocations and correcting errors in get_rows operations for Q2_K, Q4_K, and Q5_K quantizations.

These changes enhance compatibility and performance for users running llama.cpp on Intel GPUs using the SYCL backend.