"Couldn't the offsets be added to each time in the -stim_times files on a slice-by-slice basis?"
Yes. However, this would require re-writing the program to be able to do separate regressions on separate sets of voxels (in this case, from different slices). At present, the structure of the program is to pre-compute one pseudo-inverse for all voxels, then apply it to get the results. It would be a moderately drastic coding job to modify 3dDeconvolve to allow for different regressions for different voxels.