network improvements

[Jim Kingdon]

> It seems that some kind of command filter is still needed here, but
> I'm not sure that the current client-side filter in advance_into_cell
> (which only stops some double-clicks) is the best solution. Perhaps we
> can just let the host ignore invalid commands instead.

I can't help but think that something host-side would be the better
solution.  Maybe some kind of "recently killed" list if ignoring all
attacks on dead units is too drastic?  Maybe there is a simpler
or more elegant solution, but I'm not thinking of it right away.