Increased speed when building the matrix hierarchy by a factor 1.4.

- Moved expensive distance function to end candidate selection
  algorithm.
- Now using the matrix initializer without number of nonzeros saving on
  pass over all aggregates for calculating them.
- When setting up the sparsity pattern there is now need to use a set
  to prevent adding a column several times to a row as the
  CreateIterator of BCRS Matrix uses a set itself. Saved it.

[[Imported from SVN: r386]]