Add loop versioning to DetectReduction for SYCl-Bench covariance. Discussion in https://github.com/intel/llvm/pull/9187#issuecomment-1521796175.