The case against normalization in fMRI pattern analysis