llama.cpp Vulkan perf boost on Intel Arc

llama.cpp has released a Vulkan optimization targeting Intel Arc GPUs (specifically the B60 architecture). The change modifies matrix multiplication kernels to load F32 data in pairs rather than individually, addressing specific inefficiencies in Intel's memory handling. Performance tests indicate throughput increases from approximately 153 GFLOPS to 221 GFLOPS on certain tensor shapes. This update improves the viability of non-NVIDIA hardware for local LLM inference without requiring new model weights or quantization methods.