Does page lookup as needed.

The three "XXXXX"'s should be switched to false.  Doing that separately for ease
of bisecting...

Not optimized to use sortedness yet.  Also start putting in place infra
to faster reject bad data.

A version of Chandas.ttf found on some Chrome bots has 660kb of GPOS,
mostly junk.  That is causing 48 million of set->add() calls in
collect_glyphs(), which is insane.

In the upcoming commits, I'll be speeding that up by optimizing
add_sorted_array(), while also reducing work by rejecting out-of-sort
arrays quickly and propagate the rejection.

Part of https://bugs.chromium.org/p/chromium/issues/detail?id=794896

Now that pagesize is 8192, this feels better.

Fixes perf regression on some heavy fonts in Chrome's FT+HB
interaction.

See:
https://bugs.chromium.org/p/chromium/issues/detail?id=782220

More work to be done:
https://bugs.chromium.org/p/chromium/issues/detail?id=794896

To be used to optimize adding a whole bunch of (sorted) items at the same time,
as in CoverageFormat1.

It's as good as it gets, and seems to be on par with previous set implementation
in my benchmark.

Would be great if someone can double-check my bitops.

With new set implementation, this became really costy.  Optimize it.
There's more to be done, but this shaves off most of the fat.

Part of fixing https://bugs.chromium.org/p/chromium/issues/detail?id=782220

Fixes https://github.com/behdad/harfbuzz/issues/579.

Ouch!  That's what happens when one plays with increment/decrement operators!

Fixes https://github.com/behdad/harfbuzz/issues/578

Fixes https://github.com/behdad/harfbuzz/issues/577

Fixes https://github.com/behdad/harfbuzz/issues/576

c:\projects\harfbuzz\src\hb-set-private.hh(151): error C2327: 'hb_set_t::page_t::v': is not a type name, static, or enumerator [C:\projects\harfbuzz\build\harfbuzz.vcxproj]

Fixes clang "non-const reference cannot bind to vector element" error.

Fixes https://github.com/behdad/harfbuzz/pull/23

This one:

  map->mask = (1 << (next_bit + bits_needed)) - (1 << next_bit);

before the fix, the shift was done as an int, causing overflow
if it ever got to 1 << 31.  Sprinkle 'u's around.

Fixes https://bugs.chromium.org/p/chromium/issues/detail?id=634805

I experimented with replacing use of hb_set_digest_t with this new
hb_frozen_set_t, hoping to get a huge speedup for busy lookups
(like kern lookup in Roboto), but I only got 6% speendup in Roboto
and 4% in NotoNastaliqUrduDraft :(.

In preparation for fixing:
https://code.google.com/p/chromium/issues/detail?id=403594

Fixes previous commit's test.