Improve attrs hashability detection

[Tinche]

Fixes #16550

Improve hashability detection for attrs classes.

I added a new parameter to add_attribute_to_class, overwrite_existing, since I needed it.

Frozen classes remain hashable, non-frozen default to no. This can be overriden by passing in hash=True or unsafe_hash=True (new parameter, added to stubs) to define().

Inheritance-wise I think we're correct, if a non-hashable class inherits from a hashable one, it's still unhashable at runtime.

Accompanying tests.

[Tinche]

@sobolevn Does this look good to you? 🙏

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[sobolevn]

Let's wait a bit for more possible input. I will merge this in several days.

[brianmedigate]

@Tinche are you sure about the inheritance behavior? I just upgraded to mypy 1.9 and encountered this. I have a frozen=True attrs class which inherits from a non-frozen one, and it does seem to be hashable at runtime:

In [1]: import attrs

In [2]: @attrs.define
   ...: class A:
   ...:     a: int
   ...:

In [3]: @attrs.define(frozen=True)
   ...: class B(A):
   ...:     b: int
   ...:

In [4]: hash(B(a=1, b=2))
Out[4]: -1340312973636267308

In [5]: hash(A(a=1))
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[5], line 1
----> 1 hash(A(a=1))

TypeError: unhashable type: 'A'

However mypy doesn't seem to think it is:

import attrs
import functools


@attrs.define
class A:
    a: int


@attrs.define(frozen=True)
class B(A):
    b: int


b = B(a=1, b=2)


@functools.lru_cache()
def foo(b: B) -> int:
    return b.b


foo(b)

$ mypy test.py
test.py:23: error: Argument 1 to "__call__" of "_lru_cache_wrapper" has incompatible type "B"; expected "Hashable"  [arg-type]
test.py:23: note: Following member(s) of "B" have conflicts:
test.py:23: note:     __hash__: expected "Callable[[], int]", got "None"

This is using attrs 23.3.0, mypy 1.9.0.

[Tinche]

@brianmedigate I tested the other direction :/

I'll look into handling this case in a new PR.