gh-90872: Use strict unsigned long conversion for DWORD values

[jkloth]

The second part of the fix for negative timeouts.

This PR should not be backported for the reasons I state here: #32079 (comment)
To further emphasize this, see the additional change required in this PR for use of -1 to denote INFINITE.

https://bugs.python.org/issue46716

• Issue: regrtest didn't respect the timeout when running test_subprocess on AMD64 Windows11 3.x #90872

[eryksun]

Maybe we should start to gradually add test cases for _winapi in "Lib/test/test_os.py", even if it's only an internal module. For example:

@unittest.skipUnless(sys.platform == "win32", "Win32 specific tests")
class Win32APITests(unittest.TestCase):

    def test_dword_convert_negative(self):
        with self.assertRaises(ValueError):
            for h in _winapi.CreatePipe(None, -1):
                _winapi.CloseHandle(h)

    def test_dword_convert_overflow(self):
        with self.assertRaises(OverflowError):
            for h in _winapi.CreatePipe(None, 1 << 32):
                _winapi.CloseHandle(h)

[serhiy-storchaka]

The DWORD converter is also used in winreg.c and overlapped.c.

[serhiy-storchaka]

You can use DWORD(bitwise=True) instead of DWORD in places where you want to keep the old behavior.

Or you can define non-strict by default completely compatible converter:

class DWORD_converter(unsigned_long_converter):
    type = 'DWORD'

    def converter_init(self, *, bitwise: bool = True) -> None:
        super().converter_init(bitwise=bitwise)

and use DWORD(bitwise=False) if you need a strict conversion.

[github-actions]

This PR is stale because it has been open for 30 days with no activity.

[serhiy-storchaka]

winreg.c and overlapped.c were converted to the strict DWORD converter in the meantime (GH-94513 and GH-94516), so _winapi is now the only module which uses the bitwise one.

But the strict converter also rejects negative values, and this contradicts gh-132629: out of range values for B, H, I, k and K are now deprecated, while small negative values are still accepted intentionally, because passing -1 instead of 0xffffffff is convenient. The change in win_utils.py in this PR is an example -- WaitForSingleObject(handle, -1) works now and would start failing.

So the question is whether _winapi should reject negative timeouts, or keep k and only get the deprecation warning for values which are out of range for both signed and unsigned long.