My llama.cpp fork: GLM-4V vision, Qwen3-Next Delta-Net kernels, Devstral YaRN fix