Fixes type inference of type var from context

[spkersten]

For type vars, ConstraintBuilderVisitor always created a constraint in the SUPERTYPE_OF direction for type vars, irrespective of its direction field, which is wrong. This fixes both issue #360 as well as the errors reported by Travis in #443.

Currently, 15 tests are failing because the error messages changed due to the fix. Sometimes they became more accurate; sometimes they became wrong (reporting incompatible types of an assignment instead of arguments). Before I fix the messages and test cases, I'd like someone to look at the fix and confirm that it is indeed correct now.

The following example (extension of the one from #360) shows that this pull request fixes that issue and correctly reports a type error in the last line.

from typing import AnyStr

def f(x: AnyStr) -> AnyStr:
    if isinstance(x, str):
        return 'foo'
    else:
        return b'zar'

def g(s: str): pass

f('')  # OK
repr(f(''))   # OK: repr expect object and get str. Used to give an incorrect type error

g(f('string'))
g(f(b'bytes'))   # E: Argument 1 to "g" has incompatible type "bytes"; expected "str"

[spkersten]

Apparently, there is some problem with travis.sh. The failing tests are not reflected in the status.

[JukkaL]

Yeah, travis.sh is clearly broken for unit tests. Added #454. I will look into it.

[JukkaL]

travis.sh should work now.

[JukkaL]

Your fix was good, but it actually solved a different issue. There is actually a third bug (that was hidden by the bug you fixed) that causes the test failures and also seems to fix the original issue (but it just hid it). I spent a while figuring out what is going on, but I think I was able to finally fix it. I also fixed the other issue surfaced by your fix.