not tested yet

For faster calculation of closest cluster, seems to be slower in
practice

Add algorithm to balance number of points among processors

Still have to add algorithm to balance number of points. each processor
works on

Testing algorithm before adding to partitioner

PrepareReceiveInner was not called for sparse tag

Possible fix in packtagdata/unpacktagdata - mismatch between amount of
calls to MPI_Pack and MPI_Unpack

GridTool for Kmeans clustering, should eventually move to Partitioner
class

Fix ambiguity complains in autodiff with g++-4.2.1; disable extended solver tests that require matrix collections

Fix binary heap with key update in maximum product transversal
algorithm in inner solvers.

Correct global rescaling strategy in MLMPTILUC2 inner solver.