Pileoffeather

- enigne.py was wrongly using float64 for every single matrix, dtype has now been converted to float32
- float32 conversion improved execution speed by ~23% and reduced every model size in half (old float64 models should remain compatible)
- nn.backprop() has been renamed to mbgd -> mini batch gradient descent

- Refactor of the engine, computeBatch() has now been splitted into multiple functions: _run() _backpropagate() _update_weights().
- The new backpropagation() function in nn.py has been commented so that integrate new architectures is easier
- Execution time appears to have been unaffected, perhaps just slightly faster than before

- Complete syntax and file structure refactor, the module now consists of 3 files: nn.py (create, use and train neural nets), data_loader.py (for loading and converting data) and engine.py (the hidden core neural network engine wich is still < 100 lines of code).
- The new syntax makes it easier and more intuitive to create and load neural networks:
model = nn.create(*name, *layers)
model = nn.load(*name)
- The new file structure will also make it easier to update the module introducing new training architectures in the future

- Small refactor of pof.py and 1.3x speed improvement
- New pop and insert function on pof.py, enable the creation of composite neural networks

- pof.py refactor of neuralNetwork class, now has model.computeBatch() method. model.train() has been removed, pof.train(model ...) has been added as a replacement, this will allow in future to easily implement different training architectures
- pof.py 2.5x speed improvement in model.computeBatch()
- pod.py loadData() now support .gz dataset upload
- mnist example added

- Small refactor of pof.py, no speed optimization, fix relu activation function not behaving properly, pof.py reduced to < 100 lines of code.
- Added the ability to turn photos into grayscale or keep them rgb in pod.py