Added new test VF font (a subset of NotoSans-VF only containing glyphs 'A', 'Agrave' and 'T');
the VF was instanced with varLib.mutator, producing a series of full instances, which are
included as ttx files as well.
The tests run the partial instancer twice, once only instancing wght, then again for wdth,
and assert that the generated instance is identical to those.
This allows to drop an axis (aka L1 instancing) without knowing the
axis' actual default value from fvar table. One can simply call
`instantiateVariableFont` function with a `None` value for a given
axis (i.e. axis_limits={'wght': None}); the `None` value is replaced
by the axis default value as per fvar table.
The same can be done from the console script as well.
The special string literal 'None' is parsed as the Python `None`
object. E.g.:
$ fonttools varLib.instancer MyFont-VF.ttf wght=None
The code was setting GlyphClassDef.classDefs for the base font to an
empty dict then reading it from all fonts. It accidentally works when
creating variable fonts because the GlyphClassDef of the other fonts
will be used, but when mutating there is only one font.
Fix by reading the glyph classes before assigning to an empty dict.
When --recalc-bounds option is used the font extents in the head table
need to be updated, but since tables are lazy-loaded by default the
table will not be recompiled and will keep the old value. Force
recompiling the table by adding it to prune_post_subset tables, though
I’m not 100% sure this is the best approach.
Do not round them to integer, but let the caller do the rounding immediately before adding them to the default instance (or just before compiling the binary table as with glyf).
This ensures that the glyphs' left sidebearings are calculated in the same way as they were by varLib.mutator.
If we round deltas too early, then we may get off-by-one differences.
See the glyf table setCoordinates method where left sidebearings are computed.
varLib._GetCoordinates (which this method is copied from) did not return such data either.
The problem with also including component flags in the returned controls
tuple is that different masters may happen to have different component
flags (e.g. if one master has USE_MY_METRICS, another doesn't).
