Unify more functions in CUDA and OpenCL implementations of NBNXM