Add mamba ssm

[maedoc]

Adding some of the newer SSM based mechanisms would be of interest, and the following dérivation would facilitate implementation.

https://srush.github.io/annotated-mamba/hard.html

[maedoc]

"based" is an apparently simpler approach with similar performance one could use https://www.together.ai/blog/based combining

• sliding window
• linear (Taylor) attention

[maedoc]

Bitnet and handful of other tech mentioned in https://www.answer.ai/posts/2024-03-06-fsdp-qlora.html could be useful for the larger deep models.