Fix staleness on changed follow imports options

[ilevkivskyi]

This fixes the coarse grained part of #7777. I do not change semantics for daemon for two reasons:

• Performance penalty from this fix will be minor for regular incremental, but may be well visible for the daemon.
• And (more importantly) and just don't want to open this another can of worms (yet).

Some notes:

• The general logic is straightforward, invalidate dependent modules, if import options for their dependencies changed.
• It looks like the problem only affects suppressed dependencies, in all other situations we either already invalidate because something gets added to/removed from dependencies (e.g. when follow_imports = skip <-> normal), or there is no effect on the caller/importer (e.g. when follow_imports = normal <-> silent).
• This requires another entry for CacheMeta but good new is I have an idea how to get rid of another entry I added in one of recent PRs. Essentially I will need to record early (import) errors anyway to send to parallel workers after we switch to parallel parsing, and if we do this we won't need the imports_ignored entry.

[JukkaL]

Not a full review, but I measured performance on a huge internal codebase, and I couldn't see any regression. So performance impact looks minimal.

[github-actions]

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

[JukkaL]

Thanks for fixing this old but quite annoying issue! The approach looks good.