add NDTF and mLSTM/sLSTM

[AnFreTh]

This pull request introduces significant changes to the mambular/arch_utils module, including the addition of new classes, refactoring of existing code, and renaming of files. The most important changes include the introduction of the BlockDiagonal and NeuralEmbeddingTree classes, updates to the EmbeddingLayer class, and substantial modifications to the mLSTMblock and sLSTMblock classes.

New Classes

• Added BlockDiagonal class in mambular/arch_utils/layer_utils/block_diagonal.py for creating block-diagonal linear layers.
• Added NeuralEmbeddingTree class in mambular/arch_utils/layer_utils/embedding_tree.py for implementing neural decision trees with hard decision boundaries.

Updates to Existing Classes

• Updated EmbeddingLayer class to support a new embedding layer type "ndt" and integrated NeuralEmbeddingTree for numerical embeddings. [1] [2] [3]

Refactoring and File Renaming

• Renamed mambular/arch_utils/embedding_layer.py to mambular/arch_utils/layer_utils/embedding_layer.py and updated imports accordingly.
• Refactored import paths for normalization layers in multiple files to use layer_utils. [1] [2] [3]

Modifications to LSTM Blocks

• Added mLSTMblock and sLSTMblock classes in mambular/arch_utils/lstm_utils.py with detailed initialization and forward pass implementations. These blocks include convolutions, gated mechanisms, and projection layers.

Minor Code Cleanups

• Removed unnecessary blank lines in mambular/arch_utils/mamba_utils/mamba_arch.py to improve code readability. [1] [2] [3]