fix SystemError raised from PyUnicodeDecodeError_Create on PyPy 3.10

[davidhewitt]

In pydantic/pydantic-core#659 a SystemError is reported in the Pydantic test suite for PyPy 3.10:

_____________________________________________________ test_bytes_invalid_cpython ______________________________________________________

    def test_bytes_invalid_cpython():
        # PyO3/pyo3#2770 is now fixed
        s = SchemaSerializer(core_schema.bytes_schema())
    
        with pytest.raises(UnicodeDecodeError, match="'utf-8' codec can't decode byte 0x81 in position 0: invalid utf-8"):
>           s.to_python(b'\x81', mode='json')
E           SystemError: PY_SSIZE_T_CLEAN macro must be defined for '#' formats

Turns out that this is down to our PyPy implementation of PyUnicodeDecodeError_Create. Looks like there's a behavioural change in Python 3.10 to error when calling PyObject_CallFunction with a #.

There's a PY_SSIZE_T_CLEAN macro which must be defined to enable these formats, which has the effect in the C header of redirecting PyObject_CallFunction to _PyObject_CallFunction_SizeT.

I don't think we can do this automatic redirection, because it's a breaking change that leads to UB as PyObject_CallFunction on Python 3.9 and older would accept # formats but treat the length as c_int instead of Py_ssize_t. So I've opted to just expose _PyObject_CallFunction_SizeT and change our implementation to use that. I verified by hand this fixes the Pydantic test suite on PyPy 3.10, and continues to work on PyPy 3.9.

A further note - it looks like on Python 3.13 the differences between PyObject_CallFunction and _PyObject_CallFunction_SizeT will be removed (and the latter symbol removed from the header, but kept in the ABI). That's a problem for the future though 😄

[mgorny]

Thank you! I can confirm that with this patch the tests on pydantic-core pass for me.