ILU: Added host-based parallel ILU based on paper by Chow and Patel, SISC 2015.

CUDA and OpenCL backends will follow as separate commits.
Comments:
 - Convergence looks reasonable, yet additional component tests desired (triangular solve, ILU, etc.)
 - Matrix transposition only single-threaded, can be a bottleneck (less severe than for AMG, though...)
 - Best performance requires experimentation with number of sweeps and Jacobi iterations.
   Values between 1 and 4 look reasonable.