llama.cpp b11060 fixes Mamba memory layout

llama.cpp has released version b11060, addressing a memory contiguity issue in its Mamba state-space model implementation. The update ensures time-step projection inputs are contiguous and optimizes normalization steps to avoid redundant copies. Additionally, the build system now ships CUDA 12.8 and CUDA 13.3 libraries for Linux and Windows, aligning with recent NVIDIA driver updates. KleidiAI support on Apple Silicon is currently disabled in this release. The update also includes builds for ROCm 10.0 and OpenVINO.