Otherwise dumping avar table to TTX after it's been generated by
varLib could produce different float strings than when dumping the
same avar table after compiling and decompiling.
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).
