Omitting normalisation step in update_PFSMs has no effect on training tests

[rolyp]

As discussed, this seems to either be a bug in the training code, or a consequence of poorly-chosen training data. For the training tests to be useful, they have to be sensitive to breaking the training algorithm.

[tahaceritli]

See #129 for the changes. The training looks fine for the toy example; however, we still have an issue with testing real datasets. Perhaps the issue comes from the datasets used. I will try to understand this.

[tahaceritli]

See #143 for the inspection of the test datasets. I have summarized the whole inspection in https://github.com/alan-turing-institute/ptype-dmkd/blob/develop/notebooks/train-model.ipynb.