Improve performance of par_ilut and improve benchmark (kokkos#2846)

* Improve performance of par_ilut and improve benchmark

Main change: threshold_select was doing a small piece of
par_ilut on host. I assumed based on the code and ginkgo's OMP
implementation that the computational cost of this piece would be
trivial, but it was actually very expensive. Instead of relying
on std::nth_element, just make a copy of values and sort based on
the absolute value. This requires a bit more device memory, but is
20 times faster, so I think this is well worth it.

Secondary change: various improvements to the par_ilut benchmark. The
main improvements are the ability to validate the ILU via gmres+luprec
and efforts to reduce memory usage so you can benchmark larger matrices.

Signed-off-by: James Foucar <[email protected]>

* Fix complex

Signed-off-by: James Foucar <[email protected]>

* Remove resize_no_preserve, realloc works fine

Signed-off-by: James Foucar <[email protected]>

---------

Signed-off-by: James Foucar <[email protected]>

Author: jgfouca