[FEATURE] implement adam-mini

[thegodone]

Describe the feature
Based on this paper https://arxiv.org/pdf/2406.16793 Adam-mini (and AdamW-mini) converge faster and give better results for LLM and non-LLM architectures

https://github.com/zyushun/Adam-mini

[angeloskath]

Would perhaps https://github.com/stockeh/mlx-optimizers be a better project to implement this?