Remove the "last gid" hack from encoder implementation

[garretrieger]

Context: https://github.com/w3c/ift-encoder/blob/main/ift/encoder/encoder.cc#L219

When producing mixed encodings we need to ensure at each level of invalidating extension the loca/gvar/CFF/CFF2 offset arrays are large enough to accomadate any glyphs that may be added by any reachable glyph patches. This is currently handled using a crude hack of including the last gid of the original font in the initial subset (so that loca/gvar/CFF/CFF2 offset arrays are always at their maximum size).

As currently implemented this is sub optimal and will also conflict with producing a true desiccated font (#104). Instead at each level of invalidating extension we should determine what the maximum reachable glyph id is and ensure the loca/gvar/CFF/CFF2 offsets arrays are large enough to accomadate.

[garretrieger]

This will also likely require some additional functionality to be added to harfbuzz: add a way to retain glyph ids greater than the maximum glyph in the subset definition.