ExTernD: Ternary decomposition matches 4-bit quantization

ExTernD achieves accuracy comparable to q4km while using fully ternary weights and requiring no quantization-aware training. It uses slightly more VRAM than 4-bit quantization.