Test petclaw with Python kernel and multiple processes

[ketch]

I don't think it will work as coded right now. It may not be worth supporting this combination anyway, but if not we should check for it and throw a helpful exception.

[ahmadia]

I'm surprised that this doesn't work. Where are we breaking our orthogonality in assumptions between the PETSc interactions and Fortran kernels? I think if we are not willing to fully support the Python kernels we should remove them.

[ketch]

The issue was that grid.n was being used to determine loop limits. Now grid.n is the size of the whole grid, while grid.ng is the size of the processor's global vector. It is fixed.

It is quite possible that there will be features in PyClaw that we're not willing to support in PetClaw, though I would like to avoid that. PyClaw has a design goal of being used for teaching and rapid prototyping, which makes the Python kernels important.