Fixes https://github.com/behdad/harfbuzz/issues/528
"Kannada JIHVAMULIYA and UPADHMANIYA insert dotted circles"

Fixes tests/automatic-fractions.tests

If *I* get this wrong in 2017, I have no idea how others get their head around
cluster math...

Fixes tests/arabic-fallback-shaping.tests

I like to have a mode where CONTAINS_NOTDEF and CONTAINS_DOTTEDCIRCLE are not
returned.  Abused a value of -1 for that.  hb-shape now uses it.  Fixes two
of the six tests failing with --verify in test/shaping/run-tests.sh.

Disabled for now.  Will enable and fix failures after next release.

New API.

Change hb_buffer_diff to explicitly cast result of abs to unsigned when
comparing with position_fuzz to avoid unsafe signed/unsigned comparions
warnings on windows.

Based on patch from Jonathan Kew.

Needs more cleaning up and documentation.

New API:
hb_buffer_diff_flags_t
hb_buffer_diff()

We break and shape fragments and reconstruct shape result from them.
Remains to compare to original buffer.  Going to add some buffer
comparison API and use here, instead of open-coding.

Ouch!

Makes consumption easier.

We were relying on cluster merges not requiring unsafe flagging because
they get merged.  If cluster level requests no merging, then we flag
unsafe when merge would have happened.

Right now it checks for monotone cluster values.  Other checks to be added.

Fixes https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=3011

This shouldn't matter.

Always use the algorithm for non-monotone case.  It's more robust.