LLaDa and future diffusion LLM (dLLM) support

Since inceptionlabs' new diffusion model, I think we will see many open weights diffusion LLMs even just for their speed advantage, and it would be great if ollama supported them. Probably quite a lot of ollama's backend would need to be modified to allow diffusion compared to only auto regression. Here's a dLLM model that could be implemented, both in base or instruct: https://huggingface.co/GSAI-ML And here's the project page: https://github.com/ML-GSAI/LLaDA