DeepSeek-V4-Flash-0731 UD-IQ3_S: 12.5 tok/s on RTX 3090

Run via llama.cpp in text-generation-webui on an RTX 3090 24GB with 128GB DDR5 at 5600 MHz (AMD EXPO). The user had to replace text-generation-webui's bundled llama.cpp binaries as a workaround.