ILU: Added CUDA implementation of parallel Chow-Patel version.

Proper testing still pending. Two transpositions currently carried out on host,
might be a bottleneck (probably not, because there are still lots of solver iterations
to hide the overheads).