Include original mamba

[AnFreTh]

This pull request includes several updates to the Mambular package, focusing on enhancing the README documentation, adding new features, and refactoring code for better modularity and maintainability. The key changes are as follows:

Model Updates:

• Added a new model, MambAttention. Include it to the list of available models in the README (README.md).
• Provided additional installation instructions for using original mamba implementations and specified compatible torch and cuda versions (README.md).
• Included Quantile as a new distribution class for quantile regression in the README (README.md).
• Included original mamba-ssm implementation and usage of both, Mamba and Mamba2 in Mambular and MambaTab models

Feature Additions:

• Introduced the get_normalization_layer function to dynamically select normalization layers based on the configuration (mambular/arch_utils/get_norm_fn.py).
• Added a new _init_weights function for initializing weights in the mamba_utils module (mambular/arch_utils/mamba_utils/init_weights.py).

Code Refactoring:

• Refactored the Mamba class to use a configuration object for initialization, improving readability and flexibility (mambular/arch_utils/mamba_utils/mamba_arch.py).
• Updated the ResidualBlock and MambaBlock classes to include detailed docstrings and additional parameters for better clarity and functionality (mambular/arch_utils/mamba_utils/mamba_arch.py). [1] [2]
• Incorporated the pscan function from the mambapy package conditionally in the MambaBlock class to enhance selective scan sequences (mambular/arch_utils/mamba_utils/mamba_arch.py).

These changes collectively improve the usability, documentation, and maintainability of the Mambular package.