Don't let -s silence .py imports from stubs. Fix #1364.

[gvanrossum]

Implement the idea in the issue by Ben Darnell. (However, this
implementation still silences the case where main.py -> stub.pyi ->
x.py -> y.py. Hopefully that doesn't matter.)

[JukkaL]

LGTM. It's unclear to me whether it would be better to follow imports from stub files recursively -- they could bring in a lot of code. This hopefully fixes Ben Darnell's immediate issue.

[gvanrossum]

I've found a flaw in this idea.

Consider a .py file that's not mentioned on the command line but is imported both by a stub and by one of the initial sources. Now depending on processing order it may end up being referenced inconsistently -- one of the initial sources may have a silenced reference to it (i.e. it's equated to Any) but in the stub it's imported. It would be better if, given that we load the file anyway, its definitions are used everywhere, and a user might reasonably expect that behavior.

The consequences are probably minor -- maybe this use case is uncommon, and at worst it would suppress some useful errors, but those could be revealed by adding it to the command line.

@bdarnell beware of this.

[bdarnell]

Noted. I'm not currently planning to use -s any more. I'll just stubgen things as I need them or use # type: ignore.

This also ties in to #1190: I need to be able to specify that my .py files contain annotations that should be considered as valid as .pyi files. -s categorically considers .py files as inferior which undermines the rationale for inline annotations.