In both CUDA and OpenCL NBNXM kernels, we first check whether the `LJ_EWALD` macro is defined before we even consider defining it. However, it leads to an incorrect outcome only for cut-off electrostatics + Ewald LJ + no energy calculation flavor.