In various places within the generated Python/Python-ast.c, error handling generates a repr() and raises exceptions accordingly.

Currently in py3k the generated code uses PyBytes_AS_STRING() on the repr.  My understanding is that repr() should be a PyUnicodeObject, not a PyBytesObject.  This seems to be unchanged from r63682, which was a mass-change of PyString to PyBytes from 2 years ago.

This leads to a python crashing with an assertion failure:
test_compile_ast (__main__.TestSpecifics) ... python: Python/Python-ast.c:5835: obj2ast_expr: Assertion `((((((PyObject*)(tmp))->ob_type))->tp_flags & ((1L<<27))) != 0)' failed.

when invoking compile() on certain (malformed) trees of ast objects.

The attached patch:
  - extends the ast error-handling selftest with code that triggers this crash (on unpatched code)
  - fixes Parser/asdl_c.py to generate code using _PyUnicode_AS_STRING instead
  - contains the generated changes to Python/Python-ast.c

FWIW, it's not clear to what extent _PyUnicode_AS_STRING is deprecated, Include/unicodeobject.h currently has:
   *** This API is for interpreter INTERNAL USE ONLY and will likely
   *** be removed or changed for Python 3.1.
but I hope it won't be, it's far too useful.

IMO a better patch would be to use the %R format specification:
  PyErr_Format(
      PyExc_TypeError, "expected some sort of mod, but got %R", obj);
And no need to use PyObject_Repr and the temporary variable!

r86538