Support SemVer only for model versions

A model version should be defined using SemVer:

• major: A new incompatible model (a new version without default value) would be a major.
• minor: A new model that adds features but is backwards compatible is a minor (eg new column with default, removes column).
• patch: a retraining would be a patch (patch could be timestamp of the retraining for example)
• A model candidate (hyperparameter tuning) would be a metadata

So, version 1.23.20231111+candidate12 would be a candidate with id 12 created for 23 iteration of the version 1 of the model, trained on 20231111

This limitation is heavily opinionated, and could be changed later depending on user feedback.