cache singlet, non-singlet decomposition, move mellin loop inside pdfevol, use copy to speed up memory usage, keep negative branch for 1d inversion