The table structure requires that all named instances have the same
record size. Slightly refactored the compilation code to make the
logic easier to understand/maintain.
If a named instance has a PostScript name, the name is now printed as
a comment into the XML dump. Also, unit tests on `fvar` are now
exercising all code paths, both with and without PostScript names.
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).
