Don't rely on reprocessing of targets with errors in fine grained mode

[ilevkivskyi]

Consider this test case:

[case testWarnUnusedIgnoresFine]
# flags: --warn-unused-ignores
import a
[file a.py]
import b
b.x  # type: ignore
[file b.py]
# empty
[file b.py.2]
x = 1
[out]
==
a.py:2: note: unused 'type: ignore' comment

Currently it fails, presumably because we don't reprocess a.py. This however points to a larger potential problem, currently we are relying on reprocessing targets with errors. However, a better idea might be to only rely on fine grained dependencies. Of course, we can cache errors and just show them if a target had errors and was not reprocessed.

[msullivan]

I've been pondering this idea some and think it is a good idea, if we can make it work. I was not able to (in the time I've spent thinking about it) come up with an example that should rely on reprocessing errored targets in any fundamental way, but I might be missing something.

From an engineering perspective, update is currently structured around clearing the errors before looking at each file and then reprocessing errored targets in order to collect all the errors, but I think it shouldn't be too hard to just clear the errors for code we reprocess