Let the span of error types be object again.

This partially revert the change introduced in
767fce81.

Turns out AttributeNodes always start with type error, which we
don't want to propagate in some places.

Cython/Compiler/TypeInference.py

@@ -554,7 +554,8 @@ def safe_spanning_type(types, might_overflow, pos, scope):
     # to make sure everything is supported.
     elif (result_type.is_int or result_type.is_enum) and not might_overflow:
         return result_type
-    elif not result_type.can_coerce_to_pyobject(scope):
+    elif (not result_type.can_coerce_to_pyobject(scope)
+            and not result_type.is_error):
         return result_type
     return py_object_type