I have removed some redundant modulestate lookups and the testsuite now executes the decimal tests as fast as before the patch is applied. (at least for me).
May I ask how you tested the decimal performance?

Regarding the failing test:
It appears that the hackcheck() method in typeobject.c is responsible for this failure: 

static int
hackcheck(PyObject *self, setattrofunc func, char *what)
{
    PyTypeObject *type = Py_TYPE(self);
    while (type && type->tp_flags & Py_TPFLAGS_HEAPTYPE)
        type = type->tp_base;
    /* If type is NULL now, this is a really weird type.
       In the spirit of backwards compatibility (?), just shut up. */
    if (type && type->tp_setattro != func) {
        PyErr_Format(PyExc_TypeError,
                     "can't apply this %s to %s object",
                     what,
                     type->tp_name);
        return 0;
    }
    return 1;
}

As the context-type is now a heaptype the while-loop will continue to jump to the basetype of the conext-type, which is the object-type. There it ultimately fails when it compares func (context_setattr) to the tp_setattr of object.
Anyway, I do not understand the purpose of the hackcheck method, so I can not propose any kind of solution for this problem.