LU solvers scale poorly with dimensions

[guyer]

For meshes of the same number of cells, solve time should essentially scale with the number of faces per cell. If the script solverscaling.py is run with --pysparse or --trilinos the time scales roughly as 1:2:3, as expected, but the scaling is dramatically worse running with --scipy.

It turns out that's because SciPy defaults to using an LU solver. If LinearLUSolver is stipulated, then all three solver packages scale poorly:

1D pysparse:   0.383379936218
1D trilinos:   1.36989879608
1D scipy:      0.52269411087

2D pysparse:   1.19943213463
2D trilinos:   2.5458958149
2D scipy:      2.29545688629

3D pysparse:  31.7648050785
3D trilinos:  75.9411640167
3D scipy:     84.9442939758