If the `calloc` for `gsub_accels` or `gpos_accels` fails, then the
unlikely branch afterwards can be taken, which frees up the
`hb_ot_layout_t`, but since those fields can now be `nullptr`, then
we don't want to dereference them.

In `reference_table`, if the data is empty and we return early,
we still need to release the data object.

In `hb_coretext_shape`, there two edge cases where an early
return should release the attributed string.

Remove the faulty loop-detection logic. Use max-ops for breaking
out of infinite loops or heavy work.

https://twitter.com/behdadesfahbod/status/958065026262183936

Fixes https://github.com/harfbuzz/harfbuzz/issues/712

Issue originally reported at http://cppfiles.com/chromium.txt .

This fixes what appears to be a copy - paste issue which causes both
KernSubTableFormat2::rightClassTable and the 'right' argument to
KernSubTableFormat2::get_kerning to go unused.

I hope... Makes "ffi" work with Zapfino. I'm not sure if doing it correctly though...

We form the Zapfino ligature now. Yay! No further testing done.

For run-time checking.

"Hatran is written from right to left horizontally"

http://www.unicode.org/L2/L2012/12312-n4324-hatran.pdf

This ancient script was added with Unicode 8.0.

Also this spreadsheet (referenced in an inline comment in the
source code) has it as RTL: http://goo.gl/x9ilM

Fixes https://github.com/harfbuzz/harfbuzz/issues/711

Really, need better way to catch these :(.

Fixes https://oss-fuzz.com/v2/testcase-detail/5216838347653120
which is a stack overflow, not by way of infinite recursion, just
being deep.  That's disallowed anyway, so catch it as it happens,
not afterwards.