Fix crash on partially typed namespace package

[ilevkivskyi]

Fixes #16214
Closes #20658

The root cause is similar to the issue fixed by #19044, when there are multiple packages in google namespace, some with py.typed, some without. As a result, there can be a situation where we first find the untyped one, and mark ancestor (i.e. google) as suppressed, but then we find the typed one and actually add ancestor (i.e. google) to the graph. This creates a situation where we write some references to google in cache, while it may not be loaded on the warm run, thus causing a crash.

My solution is simple: un-suppress namespace ancestors if there is at least one package that needs them, so that the build graph is consistent.

Unfortunately, this uncovers another (pre-existing) bug where # type: ignores are not handled correctly on warm runs, similar to #20105. In any case, I am going to submit a separate PR for this bug (I think there may be a simple fix for it).

[github-actions]

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

[cburroughs]

(FWIW all the evidence pointed to it being the same issue as the reproduction at Qiskit/qiskit-serverless#1787, but for completeness this also resolves all AssertionError: Cannot find module for google crashes in my internal test repository.)