Minor improvements to ewald_LRcorrection

Minor improvements to ewald_LRcorrection

LJ-PME code doesn't need terms in odd powers of r or beta, so the code
might run a litte faster if expressed this way (fewer flops and fewer
registers).

Made the code that needs dr explicit by declaring dr only where it is
used.

Moved declaration of the fscal temporary to the blocks where it is
used, and commented that it is actually not the scalar force, but the
scalar force pre-multiplied by rinv. Probably that comment should
go in the generic kernels, also.

Change-Id: I052cb5a9b3bdf67a582aec0fcd99ad5da33a3b77