I have (in small parts) contributed to the python 3 conversion for numpy/scipy, and the only reason why I have done so is to be a good "python citizen", and because numpy is a major library in the python ecosystem. But I don't see any advantage in doing so. The cost is high, and the advantages near inexistent for me
Function annotations, the standard library cleanup, and other things came along for the really big ride, which was the bytes/unicode/etc ride. So, there's lots of improvements that came with 3.0 - but the nice thing is is that it continues to evolve.
We now have a New (more rational) GIL, we're probably going to have a futures package (http://www.python.org/dev/peps/pep-3148/) and so on (heck, python3 is also getting a jit). And things will continue to be added. Everyone knows and admits that the transition will take years, but now with 2.x being "done" (meaning, no more releases past 2.7) and more great work being done in 3.x, hopefully people in general will start to see the benefits.
Also, the funny thing with the GIL (even as the maintainer of multiprocessing) is that it's never bothered me. I've used hundreds of threads working with Queues/etc all the time. I actually don't use multiprocessing as much as I use plain old python threads.
For me, it's simply "I just need my dependencies" - which is a simple statement, but an incredibly hard problem.