Fix CI by avoiding setuptools 50.0

[gvanrossum]

Alternative to #9391; suggestion from pypa/setuptools#2352 (comment)

Test plan: see if all tests pass.

[gvanrossum]

Well dang! The Travis build ran and passed but the Windows builds (are those the GitHub actions?) failed, seemingly due to mypyc compilation issues. I don't know what to try next. (Could these be legitimate failures that were landed because CI was failing anyways???)

[hauntsaninja]

Probably not landed legitimate failures. The last PR landed was #9376 which kind of hinged on CI working :-)

[gvanrossum]

So maybe some different change broke the environment for Windows? Jelle's attempt to fix this (#9391) had the same Windows failures, as does harmless-looking #9387.

I note that the build logs, under "install tox", list "setuptools-50.0.3", which IIUC is the fixed version. In Jelle's PR it listed "setuptools-50.0.0". The failure in both looks the same (a whole block of tests failing, starting with testCompileMypyc. The most suspicious error seems to be LINK : fatal error LNK1104: cannot open file 'python37.lib'. Honestly I don't even know where the Windows build is configured these days. Anyway, this makes me think that we are dealing with a different failure than the setuptools thing.

[gvanrossum]

Also I'm guessing that technically excluding setuptools 50.0 isn't enough -- we need to exclude 50.0.0 through 50.0.2, since the fix happened in 5.0.3: pypa/setuptools@400b71c

[hauntsaninja]

Closing since we merged #9400