[ML] Apply consistent size penalty selecting best classification and regression model

[tveasey]

We currently apply a small penalty to prefer selecting small models when the validation loss is similar. This is based on the model size, but is also parameterised by the mean size of all models trained up to each point where we test. The mean size changes through the optimisation loop and means we don't apply a completely consistent penalty when comparing the candidate model with the current best model (whose penalty was calculated using earlier parameters). This change stores the best model size as well so we compute penalties using the same parameters.

[valeriy42]

Good catch. I have a single comment. If this was a copy/paste error, you could just fix it. If not, would you add a quick explanation, please? No need for me to review it again.