llama.cpp b9213 fixes pre-norm embedding mask initialization bug

The popular open-source project llama.cpp has released version b9213, a maintenance update focused on fixing a bug in the pre-norm embedding mask flag initialization (issue #23256). This flag controls whether the embedding layer's normalization mask is properly initialized before forward passes, affecting models that use pre-normalization (e.g., LLaMA architecture variants). The fix ensures that the mask is correctly set to avoid incorrect masking during attention computation, which could lead to subtle attention weight errors.

The release continues llama.cpp's tradition of broad platform support. Prebuilt binaries are available for macOS (Apple Silicon with optional KleidiAI acceleration, Intel), iOS via XCFramework, Linux (x64, arm64, s390x) with multiple backends including Vulkan, ROCm 7.2, OpenVINO, and SYCL (FP32/FP16). Windows builds include CPU, arm64, CUDA 12/13, Vulkan, SYCL, and HIP. Android arm64 CPU builds are also provided. This cross-platform coverage makes llama.cpp the go-to solution for running quantized LLaMA models locally on everything from servers to phones.