While looking at #140028, I found some unrelated test regressions in the
3.14 cycle. These seem to all come from #130317. From what I can tell,
that made Python more correct than it was before. According to [0], HP PA
RISC uses 1 for SNaN and thus a 0 for QNaN.
[0]: https://grouper.ieee.org/groups/1788/email/msg03272.html
* Skip sNaN's testing in 32-bit mode.
* Drop float_set_snan() helper.
* Use memcpy() workaround for sNaN's in PyFloat_Unpack4().
* Document, that sNaN's may not be preserved by PyFloat_Pack/Unpack API.
Split abstract.c and float.c tests of _testcapi into two parts:
limited C API tests in _testlimitedcapi and non-limited C API tests
in _testcapi.
Update test_bytes and test_class.
