A couple bugs joined forces to exhibit the mystery behavior of
crashes / infinite loops on OS X / wrong kerning / invalid memory
access.  Pooh!

The bugs were involved:

  - Wrong pointer math with ValueRecord in PairPosFormat1

  - Fallout from avoiding flex arrays, code not correctly updated
    to remove sizeof() usage.

We strictly never use sizeof() directly now.  And the PairPos code
is cleaned up.  Should fix them all.  Bugs are:

  Bug 605655 - Pango 1.26.2 introduces kerning bug
  Bug 611229 - Pango reads from uninitialized memory
  Bug 593240 - (pangoosx) Crash / infinite loop with Mac OS X

We were also doing wrong math converting Device adjustments to
hb_position_t.  Fallout from FreeType days.  Should shift 16, not
6.  Fixed that too.

There's still another bug: we don't sanitize Device records
referenced from value records.  Fixing that also.

Pango Bug 602408 - Invalid C++ code breaks compile with Sun C++ Compiler
(Error: A union member cannot have a user-defined assignment operator)

According to the bug:

C++ Programming Language by Bjarne Stroustrup: Chapter 10.4.12 forbids
explicitly using of union members with constructors, destructors or assignment
operations.

So we use a set() method instead of the assignment operator.  Ugly, but hey,
that's life.

The array query APIs now all do:

  - Return the total number of items
  - Take a start_offset, such that individual items can be fetched
  - The _count IN/OUT variable always has number of items written
    into the array upon return of the function

It's not part of the standard and MSVC doesn't like it.  It makes the code
a lot less elegant than it used to be, but I think it should work now.

Reported by Keith Stribley on HarfBuzz list.

Apparetly in C++, inline means a totally different thing.