add_offset was calculating DST differences wrong. It assumed to lose an hour, DST would be {-1,0,0}, in fact, DST went from {1,0,0} to {0,0,0}. add_offset now uses their differences (hence requires both as arguments)
normalize/2 was added to be able to set an arbitrary flag on a plain (i.e. native erlang datetime tuple) date. This is complicated by the fact that atoms "minimum", "max", etc. are still valid and need to fall through the normalization unaltered (for now).
