Working on the previous commit, I suddenly realised I'd made a mistake
in the design of TempSeat: you can't buffer standard output and
standard error separately and then replay them one after another,
because the interleaving of the two kinds of output might also be
significant. (Especially if the consuming Seat doesn't separate them.)
Now TempSeat has a single bufchain for all the data, paralleled by a
linked list describing each contiguous chunk of it consisting of a
single output type. So we can replay the data with both the correct
separation _and_ the correct order.
71cb9ca487