ILUT: Better performance for setup stage.

3-4 fold gains observed in solverbench.
Required some refactoring of block-ILU, partially addressing some
of the work to be done in #40.
Factors L and U are now stored separately, which implies better data locality.
Main performance gain was obtained from getting rid of std::map<> and using arrays instead.

Possible future tweaks:
 - Reduce use of std::sort()
 - Extend for better block-setup.