ml: detect tensor layout during model transfer

migan: can now run in cwhn and whcn mode (but cwhn remains faster in all cases)