Training utilities

[smguo]

Would be useful to add training utilities that

• split the data into train, validation, test for monitoring overfitting and tuning hyper parameters
• at the end of every epoch, save model if the validation loss improves. Right now the model only gets saved at the end of training.
• early stopping training if validation loss does not improve

[mattersoflight]

These are reconstruction losses observed during previous training.
"Reconstruction loss averaged over all training cell patches is 0.16 ± 0.08SD after normalization of both channels. This was further validated on the test dataset, on which model reconstruction loss is 0.18 ± 0.07SD"

[smguo]

addressed in #22