Here's a paper that includes an analysis of a similar approach that uses 5 rounds of AES under the name ARS-5: http://www.thesalmons.org/john/random123/papers/random123sc1...
I know this paper, and like this approach. Reduced-round primitives let you take advantage of existent analysis, and keep some guarantees while improving speed. In the case of AES, we know that the probability of any differential for 4 rounds is at most 2^-113; similar for linear probabilities [1]. I'm curious why ARS-4 fails Crush; I wonder if the custom key schedule is the culprit.
Additionally, a generator that passes those 200 tests is not guaranteed to be high-quality for a particular usage, which might as well be considered nothing but another statistical test. There is the famous case of the r250 generator---an additive Fibonacci generator that passed all statistical tests of the time---which turned out to have characteristics that rendered costly real-world simulations wrong [1].
That is C++, yes, but I don't think there is much of anything modern about it. Trying to save on space probably contributed to its unreadability a bit.
[1] http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.69....