Optimize xoroshiro128** generator
As mentioned in !192 (comment 13890) and nearby comments, our implementation via the VOP is quite a bit slower than the equivalent C code.
We should look into seeing if we can speed up the implementation. The C code generates reasonably compact code that is quite a bit faster than our SSE2 implementation.
The pure Lisp implementation is about 4.5 times slower than the VOP. Not great but not terrible, considering that I didn't really try too hard to optimize the speed. It would be nice if the pure Lisp version were comparable to the VOP.