Implement the roff(7) .shift and .return requests,

Implement the roff(7) .shift and .return requests,
for example used by groff_hdtbl(7) and groff_mom(7).

Also correctly interpolate arguments during nested macro execution
even after .shift and .return, implemented using a stack of argument
arrays.

Note that only read.c, but not roff.c can detect the end of a macro
execution, and the existence of .shift implies that arguments cannot
be interpolated up front, so unfortunately, this includes a partial
revert of roff.c rev. 1.209, moving argument interpolation back into
the function roff_res().