[ENH] Dense mappings should factor cost matrix by default

[pbarbarant]

As discussed with @bthirion, the cost matrix should be factored via Nystroem approximation for dense mappings as it is the case with the sparse solvers. This should greatly improve time/memory performance even for lower resolution meshes.