llama.cpp b9941 release with index-by-compile-time optimization

The llama.cpp project has released version b9941, which introduces a change to rely solely on compile times for indexing and always performs multiply/add operations.

This approach avoids relying on the compiler to optimize local memory away and is cheaper than issuing control flow statements. The release provides binaries for macOS (Apple Silicon and Intel), iOS, Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Android, Windows (CPU, OpenCL, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), and openEuler.

The update ensures consistent indexing behavior across platforms while optimizing performance by reducing control flow overhead.