Add Mamba to available LLMs

[mapmeld]

Summary

This PR adds support for Mamba models (by default mamba-2.8b) which are an alternative to attention-based LLMs.

Changes include:

• adding Mamba to the models/ directory, MambaEngine to the engines/ directory, and updates to config files
• upgrading the Transformers library dependency to v4.39, the earliest including MambaForCausalLM

Notes:

• if upgrading transformers should be avoided, the code to implement the Mamba model could be included in the models/ directory
• Mamba does support PEFT / adapters, so this could be added
• Could default to a smaller Mamba model, or selection of Mamba model sizes

Checklist

Demonstrated inference in this notebook: https://colab.research.google.com/drive/1-i4xmsyppWBdwR1qt6QN21m0uiXu0guM?usp=sharing

• Mamba is loadable with model = BaseModel.create("mamba")
• Model generates text with model.generate(...)

[MarcosRiveraMartinez]

@mapmeld Thanks for your contribution!