New I-DLM architecture breaks bottleneck in diffusion language models