llama.cpp fixes MoE dispatch inefficiency on Vulkan

llama.cpp has released an update addressing performance inefficiencies in Mixture-of-Experts (MoE) inference on Vulkan. The fix corrects the matmul tile selection logic, which previously miscalculated the number of active rows per expert, leading to significant idle time among GPU workers. This optimization is particularly relevant for large MoE models like Sarvam 30B, where dispatch overhead can dominate processing time. The update includes standard binaries for various platforms including CUDA, ROCm, and Apple Silicon.