New version of the patch, as well as one that doesn't touch Py_DECREF
but defines likely and unlikely. Then, three ceval patches that result
in small speedups (2% to 8% here), but don't play well together (neither
with the Py_DECREF one).

Since I get speedups for some changes and slowdowns when more than one
is applied, I suspect my system is hitting some limitation (register
stavation?) and contributing more noise that signal :/