Avoid parsing subtable if coverage check fails

[dfrg]

This extracts the primary coverage table (along with any suitable caches) into a field on SubtableInfo so that we can check coverage before attempting to parse the subtable. The glyph id, coverage index, coverage metadata and external cache are then passed to the subtable apply function in a new ApplyState parameter. This avoids some buffer indexing and reparsing the primary coverage table.

This attempts to clean up and unify some of the caching code paths along with laying some groundwork for future optimizations (like #260).

HR benchmarks seem to show a reasonable gain on non-Latin shaping as a bonus.

[behdad]

I see a 6% slowdown on Roboto, although speedup on heavier Arabic fonts:

Comparing before to after
Benchmark                                                                               Time             CPU      Time Old      Time New       CPU Old       CPU New
--------------------------------------------------------------------------------------------------------------------------------------------------------------------
BM_Shape/Roboto-Regular.ttf/en-thelittleprince.txt/harfrust                          +0.0655         +0.0655             7             7             7             7
BM_Shape/Roboto-Regular.ttf/en-words.txt/harfrust                                    +0.0200         +0.0205            12            12            12            12
BM_Shape/SourceSerifVariable-Roman.ttf/react-dom.txt/harfrust                        +0.0028         +0.0032            77            77            77            77
BM_Shape/NotoSansDevanagari-Regular.ttf/hi-words.txt/harfrust                        +0.0001         +0.0001            27            27            27            27
BM_Shape/Amiri-Regular.ttf/fa-thelittleprince.txt/harfrust                           -0.1877         -0.1878            39            32            39            32
BM_Shape/NotoNastaliqUrdu-Regular.ttf/fa-thelittleprince.txt/harfrust                -0.0602         -0.0597            87            82            87            82
BM_Shape/NotoNastaliqUrdu-Regular.ttf/fa-words.txt/harfrust                          -0.0594         -0.0591           103            97           103            97
BM_Shape/Gulzar-Regular.ttf/fa-thelittleprince.txt/harfrust                          -0.1554         -0.1555          2200          1858          2190          1850
BM_Shape/Gulzar-Regular.ttf/fa-words.txt/harfrust                                    -0.1409         -0.1408          1426          1225          1418          1219
BM_Shape/NotoSansDuployan-Regular.otf/duployan.txt/harfrust                          -0.1638         -0.1639          3079          2574          3065          2562
OVERALL_GEOMEAN                                                                      -0.0718         -0.0716             0             0             0             0

[behdad]

Should this be u32 at least?

[dfrg]

It's a u32 underneath. I don't have a preference on the actual type

[behdad]

Oh that's good to know. I was under the impression that it's 16bit, and hence a headache to update for beyond-64k.

[dfrg]

A while back we added GlyphId16 specifically for 16-bit glyphs and changed our common GlyphId to be 32-bit.

[dfrg]

I see a 6% slowdown on Roboto

oof, the Roboto benches are so temperamental :( I'll see if I can figure out why.

[behdad]

Patch generally looks fine to me. My only concern is that we're getting more and more distanced from HB in some of the trickiest part of the code. I'm afraid I'll have to defer to you porting any changes in subtable instantiation or caching in the future.