Fix type object signature when both __new__ and __init__ present

[msullivan]

Currently mypy will prefer __init__ to __new__ for determining
the signature of a type object if there exists any __init__
other than object's.
Instead, prefer the closest definition in the MRO, so that
subclass __new__ can override parent __init__.

Fixes #1435.

[gvanrossum]

Note that only the first example fails in CPython -- the others are OK.

The rule CPython uses is this (for a class C with superclass B, where B is not object):

• If both C.__new__ and C.__init__ refer to their versions in object, the signature is def () -> C.
• If exactly one of C.__new__ and C.__object__ refers to its object version, the other determines the signature.
• If both override their corresponding object versions, they must match -- even if they don't come from the same class (say, if one comes from C and the other from B).

[gvanrossum]

But at runtime in CPython this is an error:

$ Traceback (most recent call last):
  File "_.py", line 9, in <module>
    b = B()
TypeError: __init__() missing 1 required positional argument: 'x'

[msullivan]

Oh, hm:
The original bug example had the __new__ call return an int, which works because __init__ isn't called if it isn't a subclass.

[gvanrossum]

Let's merge this and open a new issue for the remaining nits:

• should prefer __new__ over __init__
• should check __new__ against __init__ (using the full rules I specified)