Parallel LE for GPU distributions

[kevinstratford]

Summary

If wanting Lees Edwards planes there is currently a grossly inefficient kludge which copies the distributions back and forth to perform the reprojection and the interpolation steps on the host.

The Solution

One should be able to run the LE transformation effectively on the GPU as well as the CPU.

What's required

• The existing work by Anton on the serial case
• The equivalent when MPI decomposition is present; in particular, the relevant data from the device need to be transferred via message passing.

This is #280