Looks good to me. I'm still hoping to complete #3181 at some point too once I have a chance to figure out a bytecode caching workaround. But I don't think it will conflict terribly with these changes.

Like the approach, but cannot really judge the implementation. Just to be sure something silly doesn't get passed by: numpy c code is full of Py_INCREF and Py_DECREF -- does the NPYITER stuff automatically take care of that for these purposes?

The only PyObject involved here is what's passed in (args) which is a borrowed reference (i.e. it doesn't need to be DECREF'd on exit. NpyIter is not a PyObject, so doesn't have a reference count.

I think this is pretty uncontroversial at this point. I see no reason not to merge.