llama.cpp b9934 tunes subgroup split in flash_attn_vec

The llama.cpp project has released version b9934, which includes a tune of the subgroup split (d_split) parameter within the flash_attn_vec function via ggml-webgpu.

This release provides binaries for macOS Apple Silicon and Intel, Linux Ubuntu across CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL backends, Android arm64, Windows with CPU, OpenCL Adreno, CUDA 12/13, Vulkan, OpenVINO, SYCL, and HIP support, as well as openEuler x86 and aarch64 builds.

The update also includes the standard llama.cpp UI package for cross-platform use.