perf(core): skip transparent glyph pixels

[lfkdsk]

Glyph cells are mostly empty. Rasterizing hero's text scans 10.7M glyph-cell
pixels of which 1.66M carry coverage — 84.6% of the work touches transparent
pixels. On a text-denser app (note) it is 86.8%.

This builds a sparse per-row span index when an atlas is parsed, and walks only
the covered samples during full-render glyph runs. Coverage values still live in
the bitmap; the index stores coordinates.

Measured on main

At the Vita configuration — density-2 baked atlases, scale-2 output — 180-frame
medians over 15 alternating independent-process pairs:

	dense	sparse
hero	325.255 ms	263.670 ms	−18.9%, 15/15
note	210.322 ms	141.945 ms	−32.5%, 15/15

The cost is parse-time storage: hero's d2 span indexes are 144,628 B across 10
allocations, note's 84,890 B across 8. The 180 render calls after that allocate
nothing. The feature adds 6,292 B of wasm over the dense opt-out build.

Parse happens once per guest boot — every loadFontAtlas call site is a mount
loop, a PSP/Vita pak feed, or a fresh sim/tape/web world. There is no runtime
on-demand font-slot loading, so the parse cost is never paid at frame cadence.

Where it applies

Sparse spans compile in on web/wasm, the Vita and Switch software rasterizers,
and normal hosts; they compile out on PSP, Nintendo 3DS and ESP-IDF. Real
PSP and 3DS target outputs plus a forced ESP-IDF configuration each contain zero
sparse helper symbols.

Verification

Four cargo matrices compare sparse against the dense oracle on buffers
prefilled with nonzero data, covering scales 1..4, density-2 sampling and
downsampling, four-edge and disjoint clips, negative cell origins, translucent
composition with destination reads, alpha 0, a full DrawList through the real
GLYPH_RUN dispatch, and RGBA/ARGB/RGB565 targets. Both 180-frame app replays
assert per-frame hashes.

Mutation checks: treating coverage below 64 as transparent fails 6 tests and
diverges at d2/s2 frame 0 (the threshold is chosen because baked atlases contain
no bytes 1..8, so a cov <= 1 mutant would be invisible end to end); dropping
the last source row or column at density ≠ 1 fails 2 tests each.

One documented gap: a floor-instead-of-ceil mutant in the destination-run
boundary math fails its unit test but still passes both 180-frame d2/s3 replays.
The non-integer downsample branch has no frame-level coverage today — shipped
tapes are d1/s1 and Vita is d2/s2 — so it rests on the density 1..255 × scale
1..4 unit matrix.

Suite behaviour matches clean main exactly: 574 pass, 1 skip, the same five
unit failures, and the same frame-0 divergence of the committed hero-main tape
hashes.