[codex] GPU-culled deck.gl Trace Layer

[ibgreen-openai]

Goals

• Add a WebGPU-only deck.gl trace viewer that demonstrates one GPU command graph culling 25,000 trace blocks and their Arrow-backed glyph records before indirect drawing.
• Extend the shared text and command-graph paths needed to select row-indexed glyph records without CPU readback.
• Keep mixed-case Arrow text correctly aligned and clipped in deck.gl views.

Changes

• Add the GPU-culled deck.gl trace example, website page, navigation entry, device setup, tooltips, culling statistics, and WebGPU indirect block/text layers.
• Add GPUTextSelection for graph-native glyph visibility, compaction, ordered record gathering, and indirect instance-count output.
• Add row-indexed storage text support through ArrowTextRenderer and ArrowTextLayer, including optional alignment-buffer specialization for portable WebGPU binding limits.
• Use float clip rectangles consistently across Arrow conversion and text shaders, add viewport-aware content alignment/cutoffs, and correct deck.gl text Y/baseline handling for mixed-case glyphs.
• Add unit and headless coverage for graph text selection, font-atlas baselines, source mapping, GPU text inputs, and portable storage-buffer limits.

Verification

• nvm use — passed with Node 22.22.1.
• yarn install — passed with existing peer-dependency warnings.
• yarn lint fix — passed; no fixes required.
• yarn build — passed after the final formatting and merge-resolution changes.
• yarn test — node suite passed (178 passed, 2 skipped). Headless suite completed with 1,260 passed and 25 skipped; one existing storage-polygon assertion failed locally: arrow-polygons-storage-test has drawable storage-backed instances. The identical failure and polygonViewportUniforms warning reproduce from a clean origin/master worktree, while the changed Arrow text/headless tests pass.
• yarn website:build — passed.
• (cd website && yarn build) — passed.
• Manually verified the GPU-culled trace in WebGPU at readable zoom, including aligned mixed-case storage parse, network café, and compute μtask labels.

Risks and follow-up

• The new example is WebGPU-only and intentionally requires storage buffers, compute dispatch, and indirect drawing.
• The local macOS headless polygon baseline failure above is unrelated to this diff and should be evaluated by normal CI on the PR.