In the previous table-driven implementation, client code had to
know the internals of the ‘fvar’ structure for correctly adding
variation axes to a font. In the new implementation, clients
do not have to futz around with binary offsets (which makes it
more likely that tools build correct fonts).
