Part of https://github.com/harfbuzz/harfbuzz/issues/1553

Now that we have 6daf45e0, revert cryptic hacks...

This reverts commit abd81ed4.
This reverts commit 9c6921c0.
This reverts commit d39760ca.
This reverts commit fedd8e6c.

Fixes https://github.com/harfbuzz/harfbuzz/issues/1374

Sigh.

hb-ot-kern-table.hh: In member function 'int OT::KernSubTableFormat3<KernSubTableHeader>::get_kerning(hb_codepoint_t, hb_codepoint_t) const':
hb-ot-kern-table.hh:59: error: ambiguous overload for 'operator[]' in 'kernValue[kernIndex[i]]'
hb-ot-kern-table.hh:59: note: candidates are: operator[](T*, int) <built-in>
hb-dsalgs.hh:574: note:                 Type& hb_array_t<Type>::operator[](unsigned int) const [with Type = const OT::IntType<short int, 2u>]

Geeza Pro for example, relies on that for fancy mark positioning.

Fixes https://github.com/harfbuzz/harfbuzz/issues/1405

Not sure I've marked all such structs.  To be done as we discover.

Fixes https://github.com/harfbuzz/harfbuzz/issues/1300

Untested.

That's what the spec says for Backwards flag, only applicable to
formats 1 and 4.

This, enables Format2 for kern table, which was disabled before.

CoreText doesn't, but no reason we shouldn't do.

Not used in any fonts.  Not well-specified when mixing kerning with
Cross-Stream positioning.

Tested with Waseem TTC:

$ hb-shape Waseem.ttc جحخج
[F1Jeem_R2=3@0,180+479|M1Khah_L2_R2=2@0,682+403|M1Hah_L2_R2=1@0,1184+403|I1Jeem_L2=0@0,1184+744]

Solves a mystery or two.  I'm fairly confident this is what CoreText does now.

Getting closer.  So many open questions still...