Ewald: GPU offloading

[blackwer]

This actually applies to DMK as well. There's currently no infrastructure for GPU offloading. Ewald and DMK will both benefit greatly from this, and will actually share some of the same routines. Notably offloading the all-pairs direct interactions will, after we implement #21, use the exact same code in DMK and Ewald. First step would be to add cuda detection with a cmake build option to enable it.