pyanalyze doesn't support assignment in __post_init__

[hauntsaninja]

from dataclasses import dataclass

@dataclass
class T:
    x: str

    def __post_init__(self):
        self.s = 1
        print(self.s)

@dataclass
class TT(T):
    def foo(self):
        print(self.s)

will give:

b.py.T has no attribute 's' (code: undefined_attribute)
In b.py at line 9
   6: 
   7:     def __post_init__(self):
   8:         self.s = 1
   9:         print(self.s)
                    ^
  10: 
  11: @dataclass
  12: class TT(T):

b.py.TT has no attribute 's' (code: undefined_attribute)
In b.py at line 14
  11: @dataclass
  12: class TT(T):
  13:     def foo(self):
  14:         print(self.s)
                    ^

If the subclass is not a dataclass, you don't get the second complaint

[JelleZijlstra]

Do you think __post_init__ should be special here? Pyanalyze generally assumes dataclasses have only those attributes defined as fields, because in general it's currently very hard to detect undefined attributes. So I might not want to change this until I have a more solid way to detect undefined attributes in general. That's something I've been thinking about, but it would be a big change.