Ported titanxstall test to moderngpu mergesort

[seanbaxter]

I forked your repo and submoduled moderngpu into it. It's not quite literally what you had--my mergesort reuses the input array to ping-pong the passes, so it only accepts pointers, not iterators. So I have a transform() call that combines the hash and the particleinfo into a tuple<>. I believe thrust does this internally, as I cannot see how it can ping-pong between your provided zip-iterator and a by-value store that it creates on its side.

The hash is a function of counter to keep things changing. Anyways, I've run out to 25,000 iterations now.

[seanbaxter]

One other thing is that the thrust build generates 220 STL instructions, when you look at the PTX. I think this is due to the way thrust implements tuples. Very heavy traffic into local memory doesn't mean the thing is broken but it does make me nervous. Past 44,000 iterations now.