Inverse interface freshness logic

[ilevkivskyi]

Fixes #9554

This is another case where I am surprised id didn't work like this in the first place. Right now the freshness info originates from the dependency itself (like trust me, I am fresh, whatever it means). IMO this doesn't make much sense, instead a dependent should verify whether all dependencies are the same it last seen them. On the surface the idea is simple, but there are couple tricky parts:

• This requires splitting write_cache() in two phases: first write all data files in an SCC (or at least serialize them), the write all meta files. I didn't find any elegant way to do the split, but it is probably fine, as we already have this untyped meta JSON in few places.
• I am adding plugin data (used by mypyc separate compilation currently) as part of interface hash. It is not documented whether it should be this way or not, but I would say it should, and this is essentially how mypyc expects it (group name will appear in #include <__native_group_name.h>, so it is a part of the interface). It used to work ~accidentally because we check plugin data in find_cache_meta() that is called before setting interface_hash, not in validate_meta().

cc @msullivan who wrote some of the relevant code.

[github-actions]

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

[JukkaL]

The new logic makes much more sense.