Fix cached version hierarchy comparison

[Jongmassey]

A Hierarchy contains a dict which contains collections of codes stored as sets.

We cache the hierarchy as a json string for performance reasons.

This cached version is deserialised and used for comparisons to ascertain whether the CodelistVersion it belongs to is compatible with various releases of the underlying coding system.
The standard json deserialisation uses lists not sets.
This makes the comparison of these collections of codes order-sensitive, which is not required or appropriate.

This change provides a custom deserialiser for cached hierarchies which uses sets in place of lists such that these comparisons are order-insensitive.

Fixes #2912

[evansd]

Arghh, such a subtle nasty bug! Nice work in tracking this down.

This is just a drive-by comment to say that an alternative fix would be to use pickle instead of JSON for this which will naturally preserve the native Python types. There are downsides to that of course in that it's not human readable and you cant' easily work with it outside of Python. But depending on exactly how these cached values are used it might be an option.

You could do an incremental migration here by checking the first character of the serialised data for { and using json or pickle to deserialize accordingly. And then update all the cached values to use pickle, and then remove the backwards compatibility code.

That's obviously more work than the approach you've taken here, but possibly solves the issue in a more general way.

Anyway, this was just to flag that as an option for consideration, not a recommendation.

[lucyb]

I think this is a reasonable solution. I'm still learning about this area of the codebase though and I'm wondering how Hierarchy.from_cache and Hierarchy.data_for_cache fits into this? The former uses json.loads(data), should that be updated? The latter still uses lists, would it make sense to update that to use sets (it'll be clearer even if it works okay)?