gh-151763: Fix debug assertion in unicode subtype cleanup under OOM

[zainnadeem786]

Summary

This PR addresses OOM-0030 from gh-151763.

It fixes a debug-build assertion that can occur while cleaning up a partially constructed Unicode subtype after an allocation failure.

Issue

unicode_subtype_new() allocates a non-compact Unicode subtype object and initializes its internal state before allocating the backing character buffer.

If allocating that buffer fails, the constructor cleans up using:

Py_DECREF(self);

At that point the object is valid enough to own a reference, but its internal data pointer is still NULL.

During debug builds, deallocation calls unicode_is_singleton(). For length-1 strings that helper reads the first character using:

PyUnicode_READ_CHAR(unicode, 0);

For a partially initialized non-compact Unicode object this reaches the internal assertion:

data != NULL

causing the interpreter to abort instead of propagating the pending MemoryError.

Fix

Before reading the first character, unicode_is_singleton() now detects the temporary construction state:

if (!PyUnicode_IS_COMPACT(unicode) &&
    _PyUnicode_DATA_ANY(unicode) == NULL) {
    return 0;
}

A non-compact Unicode object with a NULL data pointer cannot represent a singleton, so returning 0 is correct while avoiding the invalid debug-only access.

The constructor, ownership model, and cleanup logic remain unchanged.

Validation

Built locally:

PCbuild\build.bat -p x64 -c Debug

Reproduced the original OOM path using _testcapi.set_nomemory().

Before

The reproducer aborted with:

Assertion failed: data != NULL
Include/cpython/unicodeobject.h

After

The same allocation-failure sweep completes without triggering the assertion and correctly propagates MemoryError.

Tests

Executed focused Unicode test suites:

PCbuild\amd64\python_d.exe -m test test_str test_capi.test_unicode
PCbuild\amd64\python_d.exe -m test test_codecs

All passed.

I also verified:

git diff --check

No whitespace issues were reported.

Notes

No regression test is included.

The reproducer relies on _testcapi.set_nomemory() and allocation-failure indices that are build-sensitive, making it difficult to provide a stable CI regression test for this specific OOM path.

Addresses OOM-0030 from gh-151763.