Make TypeIndirectionVisitor depend on all classes in MROs

[msullivan]

This fixes some nasty incremental mode bugs.
Fixes #4562.

[gvanrossum]

Can you understand why the tests fail on all builds except 3.6?

[gvanrossum]

I've restarted the build, but it doesn't feel like a glitch. (There have been some glitches related to Trusty.)

[gvanrossum]

After re-running the failing build, it still fails, but now with 3 out of 4 jobs passing (previously only 3.6 passed). But 3.4 still fails, even after re-running that one. Flake? Or something else?

[msullivan]

I am deeply confused by this.

[msullivan]

I can intermittently reproduce the failure locally---it seems to genuinely be a nondeterministic failure. I am investigating.

[msullivan]

The test failures were caused by STRICT_OPTIONAL leaking out and being true for unit tests. This would happen if a test run under dmypy had strict optional on (and I added a test that did). It was nondeterministic because it only came up if the same runner that ran that test then ran the typing unit tests.

#4568 is up to fix that. I'm rebasing this diff but it needs to not land until the other has.

[ilevkivskyi]

I think you can merge both PRs when ready (if testpr is clean).

[gvanrossum]

Thanks for digging deeper!