Support snapshots

[seba-1511]

Snapshopts are unfinished experiments.

The idea is to save partial results (in a dedicated folder) that could be used to implement fancier random search algorithms. (e.g. Population Based Training.) They are also useful for very long / tedious experiments where saving intermediary results is useful.

One way to do it is to create a dedicated subfolder for these in same vein as attachments. However careful (and efficient) annotations need to be added in order to support "versioning" of snapshots. (both dumping and loading) Those should support the usual result + attachments.

Another way to do it is taking advantage of the current infrastructure, and use experiment names for versioning. This is simpler (and maybe cleaner) but it requires adding functions to modify/extend JSON summaries and saving these modified JSON summaries in a separate experiment.

[seba-1511]

Ideally, if a snapshot is extended you would be able to retrieve the history of hyperparameters that were used at each snapshot.