This does two things:
1. Intersect subsetter glyphset with the table's Coverage before
passing to ClassDef1 for subsetting. Anything that doesn't
get past Coverage wouldn't ever get to ClassDef1,
2. Never reuse class0 of ClassDef2. There's unspoken assumption
that ClassDef2's class0 is never used for actual kerning, since
that's the unbounded "every other glyph" class. Previously our
ClassDef subsetter was reusing class0 if "every other glyph"
happened to become empty because of the subset glyphset. Don't
do that for PairPos's ClassDef2. As a result of this assumption,
don't keep a PairPosClass2 subtable if only ClassDef2's class0
survived subsetting.
Would be good to add tests for both.
Related to https://github.com/harfbuzz/harfbuzz/issues/2703
8c59d4be0c