Tested, but feels fragile :(.

Assignment is invalid on invalid object.

Ugh.  Need to think of something else.

While doable with hb_deref_pointer() as well, we also would then
need to do it in a ton of places.  Not worth it / messy.

Also switch to better mixing.

No idea how to avoid dupicating code.  Was hoping hb_deref_pointer()
would do it, but looks like a pointer can't bind to a universal
reference T&&.  Humm.

Rarely used, but used, in name table or similar constructs.

Different initial and accumulator types

To be cleaned up.

Untested!

I knows...

Yeah, some things not very clear...

Fixes a bug in CoverageFormat2::serialize whereby the first range
was not serialized correctly if it consists of only a single glyph ID.
This broke shaping of U+0626 in the Arabic fallback shaper, because it
is not found in the coverage table of the 'init' and 'medi' lookups.

Also fix similar bug in ClassDefFormat2::serialize, noted during code
inspection (I haven't observed a case that was actually affected by
this, but it looks broken).

Fixes https://github.com/harfbuzz/harfbuzz/issues/1504

Untested.

coretext_aat was a temporary shaper to redirect shaping of AAT fonts
to CoreText and leaving the rest for HarfBuzz.  As HarfBuzz now supports
AAT and Chrome now actually ships that on a stable version on macOS,
we no longer care about such use-case.  If a client really wants 100%
metrics compatibility with CoreText better to use it directly or through
our API.  Replicating the same behavior still is possible using
hb_shape_full, something we don't care or like to offer anymore.

Fixes https://github.com/harfbuzz/harfbuzz/issues/1478

Should be improved for hb_bytes_t.