gh-158841: _PyBytes_Resize: for bytes subclasses, return new bytes

[encukou]

This reuses the test from #158846, but makes the change _PyBytes_Resize by avoiding the optimization for subclasses, and always returning exact bytes.

• Issue: _pickle crashes when read() returns a bytes subclass for large payloads #158841

[vstinner]

I don't think that it's correct to convert a bytes subclass to bytes. I would prefer to raise TypeError and return NULL (set *bytes to NULL). It sounds weird to that that in the caller, _PyBytes_Resize() changes the argument type.

The issue has been seen in pickle.load() with special code designed to trigger the code. IMO pickle should be fixed to never pass bytes to _PyBytes_Resize(). Did the bug have been seen elsewhere so far?

In the current code, _PyBytes_Resize() calls PyObject_Realloc(v, PyBytesObject_SIZE + newsize) to resize the bytes subclass in-place. The allocated size is wrong, a subclass needs more bytes (see bytes_subtype_new()). So the current code does fail / crash. It confirms that no code can rely on passing bytes subclass to _PyBytes_Resize() since it doesn't work (fail / crash).