Fix __Pyx_GetExcValue to avoid (possibly) running Python code while the thread state is inconsistent

For example, in the SAGE source we have

1158 PyInt_FromLong(0)
 776 PyInt_FromLong(1)
 258 PyInt_FromLong(2)
  33 PyInt_FromLong(3)
  21 PyInt_FromLong(10)

and a thousand or so others... Who knows how many of these are in loops too.

This is what the interpreter does, and allows one to
get at the actual object (rather than just its truth
value).

gcc complained that a variable might be used uninitialized, which is
true. If NULL is passed, all code gets executed with an
uninitialized variable which is an error in almost every case.
Since python itself never passes NULL, only a real error in
other c code could trigger the assert.

scope-entry as a variable. This means that
analyse_attribute expects to find a pointer to
a member of the c-object struct.
This is not the case for members, so we need a
special case.

Fixes

cdef int a, b
cdef int *foo = &a if a else &b

so the generated c files don't depend on the version of python used to run Cython

slots for new pyton versions are not emitted
for older versions. This stops compiler warnings.

The issue here (as pointed out by Thomas Hunger) is that once the CClass
body gets executed the class is already constructed, and its methods
are actual methods not functions (as in Python).

The CClassScope now returns a utility function on lookup of "classmethod"
that creates a class method out of a method (rather than out of a function).
When added to the type dictionary, the result is exactly the same as setting
the METH_CLASS flag.