Fix #996

[refi64]

@bcdarwin @gvanrossum Please try and let me know if it works!

[gvanrossum]

Sadly doesn't look like it fixes #991 for me. So maybe the two issues are not the same.

[bcdarwin]

This doesn't seem to crash any more, although it's still possible to omit the type arguments entirely (i.e., annotate a variable with type C rather than C[int]) and silently pass the checker -- not sure that's intended.

[JukkaL]

Ah, a test case seems broken. Anyway, this will fix the crash but there could be other similar issues that this won't catch :-(

I believe that the root cause of the crash is that recently I changed mypy to type check code even if there are semantic errors. However, the type checker assumes certain invariants to hold after semantic analysis, such as the number of type arguments being correct, and this no longer is necessarily the case.

I'm thinking about a more general fix where we replace bad types with Any or something during semantic analysis, so that the type checker can still assume that things are correct. This would make the type checker simpler (actually no more complex than it is right now).

[JukkaL]

I pushed a hopefully more general fix to this issue, so I'm closing this PR. @kirbyfan64 thanks for the original PR!