llama.cpp b9931 speeds MoE models with OpenCL tile-skip optimization

The llama.cpp team has released b9931, which brings a significant OpenCL optimization for Mixture-of-Experts (MoE) models. The key change is a 'ragged-tile MoE prefill FP16 GEMM' optimization that skips padded expert tiles. MoE prefill groups tokens into tiles of 32; when the upper 16 slots are all padding (router index 0xFFFFFFFF), the second half of the dot product is skipped. This cuts compute almost in half for low token-per-expert loads, which is common during inference.

Further refinement adds quarter-granularity skip: instead of skipping only half-tiles, the kernel now divides tiles into 8-column groups and skips each empty trailing group independently. This is controlled via the GGML_OPENCL_MOE_RAGGED_GRAN environment variable (options: 8 for quarter, 16 for half, 32 for off; default is 8). The optimization is applied to all eight *_f32_ns MoE GEMM variants and produces byte-identical results, ensuring no accuracy loss. The release also packages builds for multiple platforms including macOS (Apple Silicon, Intel), Linux (x64, arm64, s390x, Vulkan, ROCm, OpenVINO, SYCL), Windows (x64, arm64, CUDA, Vulkan, OpenVINO, HIP), Android (arm64 CPU), and iOS (XCFramework).