llama.cpp b9985 adds fused top-k MoE for SYCL

The llama.cpp project released version b9985, introducing a fused top-k Mixture of Experts (MoE) implementation for SYCL. This update addresses review feedback by moving the fusion dispatch logic and exposing it via the GGML_SYCL_ENABLE_FUSION environment variable.

• Added fused top-k MoE support for SYCL backends.
• Configured fusion dispatch control through the GGML_SYCL_ENABLE_FUSION environment variable.
• Updated startup logs to print the GGML_SYCL_ENABLE_FUSION status like other environment variables.
• Disabled KleidiAI support for macOS Apple Silicon builds.

The release provides binaries for macOS, Linux, Windows, Android, and openEuler across CPU, GPU (CUDA, Vulkan, ROCm, OpenCL), and AI accelerator backends.