Conformance: drop questionable non-data-descriptor dataclass tests

[ashishpatel26]

Closes #2259.

dataclasses_descriptors.py contained a DC2 section testing a non-data
descriptor (Desc2, only __get__) used as a dataclass field. Its assertions
do not match the runtime:

• DC2.x / DC2.y raise AttributeError — there is no Desc2 object in the
  class namespace (those fields have no default).
• dc2.x / dc2.y / dc2.z are the Desc2 instances stored as instance
  attributes; a non-data descriptor is shadowed by the instance dict, so
  __get__ never runs and the values are not int / str.

As noted in the issue, this behavior is under-specified and type checkers
disagree (pyrefly rejects it; mypy and pyright accept the incorrect
assertions). This removes the DC2 cases, keeping the well-defined data
descriptor (DC1) test, and rescoring:

• mypy, pycroscope, ty: were Partial only because of the spurious DC2
  assertions → now Pass.
• pyright, zuban: unchanged (Pass).
• pyrefly: still Partial — it independently rejects DC1's data
  descriptor whose __get__/__set__ types differ; note updated accordingly.

The non-data-descriptor-in-dataclass behavior can be re-added once it is
specified.

[davidhalter]

What - in your opinion - would be the proper solution for these cases? I generally want to avoid removing tests and rather expand the docs or correct the tests. There are a lot of cases where the spec is lacking and I want to avoid removing them all.

In my opinion it is very valuable to standardize descriptors on Dataclasses. I'm not particularly invested in how they are standardized, but I would like to have that to be one specific way.

[JelleZijlstra]

As I wrote in the issue, pyrefly's behavior seemed the most correct to me. I'm not sure this behavior needs to be specified in the typing spec in much detail since what I view as the correct behavior primarily reflects the runtime.