Adding support for Mamba version 2

[Goekdeniz-Guelmez]

No description provided.

[awni]

Is this working now? Have you tested it recently?

[Goekdeniz-Guelmez]

Yes, but still the same gibberish output from my last test, I still couldn't find it, yet.

[Goekdeniz-Guelmez]

using the original models from state-space, I get decent outputs, but its just mamba codestral thats not working:

python -m mlx_lm.generate --model mlx-community/mamba2-130m-8bit --prompt "Rene Descartes was" -m 120
Fetching 6 files: 100%|████████████████████████████████████████████████████████████████████████████████████| 6/6 [00:00<00:00, 57195.05it/s]
==========
not only has a great deal to offer, but it also has a lot of value.

The company has a great deal to offer, but it also has a lot of value.

The company has a great deal to offer, but it also has a lot of value.

The company has a great deal to offer, but it also has a lot of value.

The company has a great deal to offer, but it also has a lot of value.

The company has a great deal to offer, but it also has a lot of value.


==========
Prompt: 6 tokens, 139.904 tokens-per-sec
Generation: 120 tokens, 62.461 tokens-per-sec
Peak memory: 0.823 GB