Do not cache module lookup results that may become invalid in future

[sterliakov]

Fixes #19037. See my analysis in #19037 for motivation. As a side note, with this patch the second run behaves as expected (without it mypy tries to reanalyze a lot of modules during the 2nd run, now it only rechecks the inputs with errors).

[github-actions]

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

[sterliakov]

Okay, this is not the same cause as in #10360 and #16149 - they are surprisingly unrelated.

[sterliakov]

I don't see how to deal with the whole problem yet, but this looks like a sensible first step. The issue ultimately is that we treat namespace ancestors of typed packages as if they were typed. We only start doing that once we are actually aware of some nested typed subpackages, and this causes issues I linked above. At least removing the cache dependency is good IMO, but I'll need to look deeper to find a complete solution.

[JukkaL]

This looks good to me, thanks! Incremental mode bugs are usually pretty bad, and if users turn off incremental mode as a result, they will often get quite poor performance.