Add opt-in glyph caching to vello_cpu and vello_hybrid backends

[nicoburns]

Summary

Exposes Vello's (experimental) glyph atlas cache in anyrender_vello_hybrid and anyrender_vello_cpu. Off by default.

• Runtime flags only (no cargo feature):
  • hybrid: VelloHybridRendererOptions::glyph_caching(bool) (+ public field), VelloHybridScenePainter::with_glyph_caching(bool), WebGlScenePainter::with_glyph_caching(bool)
  • cpu: VelloCpuScenePainter::set_glyph_caching(bool), VelloCpuImageRenderer::set_glyph_caching(bool). VelloCpuWindowRenderer is a generic softbuffer/pixels wrapper with no options hook, so caching can't be enabled there.
• Implementation is just .atlas_cache(enabled) on the glyph_run builder, plus a guard (below).

Note: VelloHybridRendererOptions gains a public field, which is breaking for anyone building it with a struct literal.

Skewed/rotated runs bypass the cache

glifo 0.4 only validates the transform when inserting a glyph into the atlas (supports_atlas_caching), not on lookup, and the cache key doesn't include skew/rotation. So once an upright glyph is cached, a skewed run (e.g. synthetic italics) with the same font/size is drawn from the upright entry. supports_glyph_caching therefore disables the atlas for any run whose transform * glyph_transform isn't a positive uniform scale + translation.

tests/glyph_caching.rs (vello_cpu) covers this: it fails (2525 px differing by >64) with the guard removed.

Performance

Blitz paint_bench on https://en.wikipedia.org/wiki/Barack_Obama, 1366x768 @2x, 100 frames per run, 3 interleaved runs per cell. Times are the median µs/frame. "scroll 40" moves the viewport 40 CSS px every frame.

Correction: an earlier version of this description had "scrolling" rows that were actually static frames (the benchmark's scroll range was computed as 0). The tables below are from reruns with that fixed.

macOS, Metal (macOS 26.5.2 VM, "Apple M4 Pro (Virtual)", 12 cores, wgpu adapter Apple Paravirtual device). This is a GPU-backed Metal device but paravirtualised, not bare metal. Values are run1 / run2 / run3.

backend	scroll	cache	encode (paint_scene)	rasterize	total	first frame
hybrid	0	off	3201 / 3197 / 3234	3373 / 3654 / 3496	6593 / 6918 / 6756	219–239ms
hybrid	0	on	1143 / 1149 / 1123	3421 / 3447 / 3000	4546 / 4629 / 4121	223–272ms
hybrid	40	off	3671 / 3755 / 3621	2391 / 2514 / 2169	6091 / 6309 / 5808	214–241ms
hybrid	40	on	1106 / 1144 / 1117	2282 / 2283 / 2155	3527 / 3545 / 3367	218–229ms
cpu	0	off	3503 / 3383 / 3140	1124 / 1188 / 1052	4865 / 4902 / 4463	6–11ms
cpu	0	on	1179 / 1227 / 1201	3728 / 3801 / 3630	4942 / 5053 / 4855	31–42ms
cpu	40	off	5867 / 5490 / 4977	907 / 808 / 915	7345 / 7310 / 6417	6–8ms
cpu	40	on	1540 / 1581 / 1458	7201 / 7716 / 7619	8731 / 8899 / 9084	18–19ms

Linux, no GPU (8 cores; "hybrid" runs on Mesa software Vulkan, so its rasterize column is CPU-emulated and not representative). Values are min–max over the 3 runs.

backend	scroll	cache	encode (paint_scene)	rasterize	total	first frame
cpu	0	off	2149–2218	1370–1389	3504–3592	6–7ms
cpu	0	on	1474–1503	2542–3759	4053–5250	42–43ms
cpu	40	off	3427–3453	801–897	4239–4544	7–8ms
cpu	40	on	1819–1865	6060–7611	8145–9821	43–60ms
hybrid	0	off	4678–4729	38481–38883	43206–43681	111–115ms
hybrid	0	on	1936–1957	39149–40626	41207–42564	114–124ms
hybrid	40	off	5460–5490	32205–32620	37936–38325	112–113ms
hybrid	40	on	1692–1727	30731–31282	32475–32830	111–116ms

• vello_gpu (hybrid): encode time drops ~60–70% on both machines, rasterize is unchanged within noise. On Metal that is a net ~-33% per frame static and ~-42% scrolling. No consistent change to the first frame.
• vello_cpu: not a win. The work moves from encode to rasterize and grows: static is a wash on macOS and slower on Linux; scrolling is ~20% slower on macOS and ~2x slower on Linux. The first frame is several times slower (atlas population).
• Output with cache on vs off: ~3–4% of pixels differ, all antialiasing-level (max channel delta 59–60).

Link to Devin session: https://dioxus.staging.devinenterprise.com/sessions/9fbb960ba27249bca490d138236e0bd1
Open in Devin Desktop: https://dioxus.staging.devinenterprise.com/desktop/session/9fbb960ba27249bca490d138236e0bd1?variant=devin-insiders
Requested by: @nicoburns

[staging-devin-ai-integration]

I'll fix CI failures and address comments from users with write access that start with 'Devin'.

• Disable automatic comment, CI, and merge conflict monitoring