Which dim of down_proj the trellis runs along: out codes the ROTATED dim
(the fold's R1 acts on down's output) and decodes through a column gather;
in codes the un-rotated intermediate dim so down decodes with the same
plain reduce kernel as gate/up.
Which dim of down_proj the trellis runs along:
outcodes the ROTATED dim (the fold's R1 acts on down's output) and decodes through a column gather;incodes the un-rotated intermediate dim so down decodes with the same plain reduce kernel as gate/up.