PME GPU FEP sometimes does not calculate splines correctly
In pme_spline_and_spread_kernel
with two grids, the splines are computed only once. However in the spline computation there is a check for whether atomCharge
is zero, and the spline computation is omitted if so. Since that charge can be different in the two FEP states, the splines should be computed differently. Simplest is probably to compute the spline for all atoms when doing FEP, but that means we need to template the spline computation accordingly.