Improve handling of "lambda: None"

[rwbarton]

"lambda: None" might be used either where a function returning an
Optional[T] is expected, or where a function not returning a value is
expected. Previously it would always treated as the former; now the
context is used to decide.

When no context is available, the lambda is now treated as a function
not returning a value. This seems more consistent with the fact that
you can't define a variable 'a = None' without a type annotation or
declare a function to return NoneTyp. (If really desperate for the
other meaning of "lambda: None", you can write a generic function.)

Fixes #1425.

[gvanrossum]

Maybe @ddfisher can review/merge this? It might affect his work on Optional[T].

[rwbarton]

Yeah, let's just hold off on this until the strict optional checking changes land.

[ddfisher]

That'd be great, thanks!

[JukkaL]

@ddfisher This is ready to review now?

[ddfisher]

LGTM!

[rwbarton]

After rebasing this commit, one of the new strict optional checking tests fails:

[case testLambdaReturningNone]
f = lambda: None
x = f()

This tests the old behavior (which didn't change with strict optional checking) of returning NoneTyp here rather than Void. Should we update that test, or do we want that old behavior?

I don't have a strong opinion here, though I do feel that maintaining this NoneTyp/Void distinction in the type system is not worth it if we're not going to give the user a way to refer to both types... maybe discussion should move back to #1425.

[rwbarton]

I really don't care much about the behavior here with no context, while the situation with a context arises frequently (see the just-submitted duplicate #1710), so I'm going to leave the no-context behavior as is. Pushed a new version of this commit with a rationalization for that behavior :)

[ddfisher]

Sounds good to me. I also don't have a strong opinion here, aside from agreeing that we should kill Void.