Design a solution for caching downloads of `$refs` in order to improve performance in cases with many remote refs

[sirosen]

Original use-case sourced from this PR: #451

The current caching capability significantly improves runtimes for remote schemas when there is a single remote file to download, but does nothing to improve the case where there are refs to resolve. Refs are cached in-memory by referencing, but discarded between runs.

For faster runs, check-jsonschema should cache resolved refs on disk as well.

Some basic requirements:

• this must respect the --no-cache setting
  • probably the same object which is used for fetching remote schemas should be passed to the ref resolver
• filenames must be chosen such that there are no conflicts between different schemas (users won't be able to control filenames)
• if the new file-and-dir layout for these data conflicts with the existing cache dir layout, that needs resolution
  • ideal: design a strategy to migrate cache data for the next 1-2 calendar years
  • acceptable: ignore old cache data, provide a changelog note on how to clean it up
• the behavior here need to be tested

[!NOTE] A friend of mine suggested putting cache data into a DB (e.g. sqlite) when we talked about this, so that it could be annotated with richer metadata and structure. Although that might be a good idea longer term, I don't want to reach for that quite yet -- I think this can be solved with a good dir structure for now.

Here's one initial idea, for evaluation:

• each $ref is canonically named {md5 of the absolute URI}.json
• in the ~/.cache/check_jsonschema/ dir, add a dir named refs/ (the schemas are in a dir named downloads/, which now seems like a suboptimal name but will suffice)
• ref resolution stores resolved refs in the refs/ dir