Use fixed format for cache metas

[ilevkivskyi]

This makes cache meta files ~1.5x smaller. (I hoped together with previous diff it will give us 4x, but it is more like 3x). Implementation is mostly straightforward, here are some comments:

• I make CacheMeta a regular class, it doesn't need to be immutable IMO (since we actually mutate it in few places).
• I remove all uses of untyped dicts in favour of CacheMeta (note this might make JSON format slightly slower actually, but difference is below noise level)
• Instead of manually checking some individual keys, I use blanket try/except (KeyError, ValueError) when deserializing metas.
• In one place (where we update meta file after read), I update the loaded view to match the updated file 1:1. This should be more robust.
• I still use JSON dumps for options and plugins snapshots. Serializing these using FF is tricky (and will be simpler with type tags).
• I rename data_json/meta_json paths to data_file/meta_file everywhere.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[JukkaL]

Nice! Did you see any measurable performance differences?

[ilevkivskyi]

I didn't do any precise measurements, on average of three warm runs import torch was ~5% faster. I will try making more precise measurements.

[ilevkivskyi]

OK, after more careful measurements it is more like 4%, but still not nothing. Btw I also compared this PR vs 1.18.2, and warm import torch is ~25% faster.