Model featuring FF algorithm and architecture

Subtasks

[x] Utility to convert MNIST dataset from csv to tinn format @FedericoRubbi
[x] Negative sample generation. Check dataset shuffle @Congiuntivo
[x] Change NN input type from to <sample, label> @Congiuntivo
[x] Loss function accepting two passes instead of a single pass @Congiuntivo
[x] Feed two passes for training instead of a single pass @FedericoRubbi
[x] Fix gradient calculation for the implemented loss function @FedericoRubbi
[x] Output normalization in between layers @FedericoRubbi
[x] Train loop taking num layers as parameter @FedericoRubbi
[x] Inference loop @Congiuntivo
[x] Implement digit dataset (8x8 digit dataset) @Congiuntivo
[x] Implement ADAM optimizer @Congiuntivo
[x] Test model with many layers and with MNIST @Congiuntivo
[x] Implement batched training @FedericoRubbi
[x] Implement dataset test/train split @FedericoRubbi
[x] Add model utils save_model and load_model to write and read from a file the architecture (number of layers) and all the weights @FedericoRubbi