Added NgramModel - Githubissues

Added NgramModel which is the implementation of a naive Ngram model in PyTorch, using Laplace Smoothing
Added a method to generate a pseudo dataset from the target model (which is then used to fit the NgramModel).
- Some scripts are provided to executes the file
Including the Ngram model in train.py and eval.py.
- Train: not really training but allows to fit the model on the pseudo dataset or directly the given dataset and then save the ngrams
- Eval: load the model to then use speculative decoding

TBD

Implement the fit of Ngram on the dataset and not the pseudo dataset
Fix the saving method of the Ngram
Test if the eval.py file works with this model (it should since I get the input as the other models and return the logits and proba)