Incomplete LDL Decomposition

@martin.nolte implemented a ildl decomposition. this MR ports the ildl decomposition and preconditioner to the master branch.

Could you briefly comment on the relaxation factor? From my understanding incomplete LDL can be made incomplete in a few different ways: what the relaxation factor means is not clear to me from the name and I wonder what the advantages and disadvantages of this criterion are.

The relaxation factor works the same way it does work for the SeqIlU0 andSeqILUn implementations. It is just a factor which the back-solve result is multiplied with. Otherwise the ILDL is incomplete in the way that no fill-in is allowed similar to ILU0.

mentioned in merge request dune-fem/dune-fem!164 (merged)

mentioned in commit dune-fem/dune-fem@f271aec5

mentioned in commit a072b840

merged

Is it guaranteed that the implementation preserves non-negative diagonal entries? Then we would finally have a simple spd preconditioner for CG and should document this.

@carsten.graeser: Your use case was the motivation for writing this in the first place. If the matrix is dense, this assumption obviously holds. For the sparse case it seems to work very well. However, I have not found corresponding theoretical results.

mentioned in merge request !513 (merged)