Hox

- Improvement in execution time during training, mnist 1 epoch 96% accuracy on a 12600k went from 4.25 seconds to 3.5 seconds

Heavy refactor of the entire codebase
- utils removed
- training loop moved from utils to hox.py
- hox.py is now all you need to train a neural network
- Different layers and optimizers have been removed/hidden. Dense, Optimizer, SGD have been removed, the optimizer has been integrated in the Model and Layer class
- Training speed is equal to the previous version, possibly slightly faster