SYCL: Reduce the number of atomic ops in NBNXM fShift calculation