Make GICP use CorrespondenceEstimation

[mvieth]

The main benefit is that CorrespondenceEstimation is parallelized, so GICP can benefit from that. But even with 1 thread, the new code is slightly faster due to the improved way of computing the mahalanobis matrix (invert3x3SymMatrix uses the fact that the matrix is symmetric and is thus faster than the general-case .inverse()).
Otherwise, the behavior is unchanged.

[larshg]

This is a subset of #5287 ?

[mvieth]

This is a subset of #5287 ?

Oh, right, I forgot about that pull request. Comparing the two, there are definitely some similarities (replacing GICP's for-loop + searchForNeighbors by determineCorrespondences of a correspondence estimation class), but also some differences, for example my improvement of the mahalanobis matrix computation and explicitly activating parallel correspondence estimation via setNumberOfThreads. In addition, I think the changes regarding the correspondence criteria in #5287 would need to be carefully examined: does it make sense to add a new class GICPConvergenceCriteria? Could DefaultConvergenceCriteria not be used? Overall, I think that #5287 applies a lot of changes at once and would require quite some time for reviewing.