llama.cpp b9290 adds SYCL Level Zero detection for GPU optimizations

ggml-org has released llama.cpp b9290, a maintenance update to the widely-used open-source C/C++ implementation for running large language models locally. The headline feature is a centralization of SYCL Level Zero detection within the ggml_sycl_init function. Previously, Level Zero initialization was handled inconsistently across different SYCL backends, leading to potential detection failures on Intel GPUs and other Level Zero-compatible devices. This patch streams that logic into a single call, using the same error messages and fallback behavior across all SYCL configurations, making GPU detection more robust and reducing developer friction.

Beyond the SYCL fix, b9290 provides a comprehensive set of prebuilt binaries for virtually every common platform: macOS (Apple Silicon ARM64 with and without KleidiAI acceleration, Intel x64, iOS XCFramework), Linux (CPU-only on x64, ARM64, s390x; Vulkan-accelerated; ROCm 7.2 for AMD GPUs; OpenVINO; SYCL FP32 and FP16), Android (ARM64 CPU), and Windows (x64 and ARM64 CPU; CUDA 12.4 and 13.1 DLLs; Vulkan; SYCL; HIP for AMD). This broad coverage ensures that users can leverage llama.cpp for local inference on almost any hardware, from personal laptops to dedicated AI servers. The release also includes updated dependencies and build scripts, making it easier for developers to compile custom versions. As local LLM deployment becomes more critical for privacy and cost-sensitive applications, such incremental but impactful optimizations are essential for keeping llama.cpp competitive with proprietary solutions.