Support for nanogpt (and gpt-j)

[uogbuji]

I'm hoping to use MLX-LM to try out inference with this new LM specialized for Nigerian languages. It's based on Karpathy's nanogpt (they say it's inspired by GPT-J's architecture—I suspect inferencing over base nanogpt would work). As expected looking at the supported model types list, core MLX doesn't support that. I also see the separate nanoGPT_mlx, but looks like work really slowed down on that.

Supporting a Nigerian language LLM would be a really nice nod to the global south (it includes Pidgin, which is widely spoken in Africa). I think it would also bring a lot of interest to MLX for fast inference as this model develops.