llama.cpp SYCL Performance Boost for Q4_K FFN

llama.cpp introduces a SYCL optimization that fuses operations for Q4_K dense FFN layers, leading to significant performance improvements. Benchmarks show up to +2.8% for single-batch inference and up to +12.4% for batched inference on specific models like qwen2.5-3B and gemma-2-2b-it.