Shadowing and unshadowing a builtin makes pyflakes emit a F821

Consider this silly code:

int = 42
del int

x = int('42')
print(x)

This is perfectly valid Python: if you run it, it'll print 42.

flake8 emits a F821: undefined name 'int' on the penultimate line. I think it's because of the del int.

For a real-world example of this situation, consider a project that uses gettext and installs _ in builtins, but doesn't do that early enough to be available at import time of every module. Here's how linkchecker copes with that situation:

_ = lambda x: x
Fields = dict(
    realurl=_("Real URL"),
    cachekey=_("Cache key"),
...
    level=_("Level"),
    modified=_("Modified"),
)
del _

and this then causes flake8 to complain about _() being used elsewhere in the file:

./linkcheck/logger/__init__.py:278:19: F821 undefined name '_'
./linkcheck/logger/__init__.py:340:16: F821 undefined name '_'
./linkcheck/logger/__init__.py:377:22: F821 undefined name '_'
./linkcheck/logger/__init__.py:380:22: F821 undefined name '_'
./linkcheck/logger/__init__.py:382:22: F821 undefined name '_'
./linkcheck/logger/__init__.py:390:22: F821 undefined name '_'

flake8 --version reports

3.8.2 (flake8_2020: 1.6.0, mccabe: 0.6.1, pycodestyle: 2.6.0, pyflakes: 2.2.0) CPython 3.8.2 on Linux

Start by running the provided shadowing and unshadowing example with pyflakes or flake8, then compare it with the linkcheck example. Trace the undefined-name handling around del and add a regression test showing that valid builtin or gettext uses after deletion do not emit F821; the test should still catch genuinely undefined names.