feat(examples): add GPU-native deck.gl graph exploration

[ibgreen-openai]

Goals

• Add a genuinely usable, optional deck.gl graph explorer without introducing deck.gl into any shipped luGraph or experimental-module dependency.
• Render, explore, and pick the same GPU-resident graph allocations already produced by luGraph analytics and progressive layout.

Changes

• Add an unpackaged examples/deck/lugraph-explorer integration containing a real deck.gl Effect, an OrthographicView application, and custom GPU-resident node and edge Layers.
• Encode GPU CSR topology, PageRank, weak components, breadth-first neighborhood selection, and progressive force integration into deck.gl's own frame command encoder; deck.gl retains submission ownership.
• Bind live float32x2 layout positions directly as instanced vertex attributes and consume PageRank, component IDs, distances, and selection masks as GPU storage.
• Create one edge layer per original nonempty source/target chunk while preserving empty intermediate batches and stable source vertex IDs.
• Support native asynchronous WebGPU picking, neighborhood selection, drag-to-pin, reset, release, zoom, dynamic traversal depth, and optional caller-supplied datasets without reading graph columns back.
• Correct actual WebGPU layer instance counts, reserved WGSL identifiers, and OpenGL-to-WebGPU clip-depth conversion revealed by real-device tests.
• Add a narrowly scoped, one-call Deck 9.3 shader-assembler compatibility bridge that restores the original static method before forwarding the initialized device's actual shading language.
• Register a lazy WebGPU-only gallery route with curated metadata, an existing thumbnail, and zero-copy optional integration documentation.
• Add package-isolation/catalog Node regressions and genuine Chromium/WebGPU integration coverage. The Deck presentation regression uses a real test-owned GPU framebuffer, following the existing repository workaround for expired SwiftShader external presentation surfaces, while still asserting actual Layer drawing, queue submission, and native stable-ID GPU picking.

Verification

• nvm use: passed; Node v22.22.1.
• yarn install: attempted; the existing unpublished @vis.gl/dev-tools@2.0.0-alpha.4 dependency returns HTTP 403. Used an independent APFS clone of the existing compatible dependency tree without lockfile or dependency changes.
• yarn lint fix: passed after final changes.
• yarn build: passed after final changes across every package.
• yarn test-node: 1,026 passed, 1 skipped.
• yarn test: 1,026 Node tests and 1,709 Chromium/WebGPU tests passed; 25 browser tests skipped by their existing capability guards.
• Focused genuine, unshimmed Chromium/WebGPU Deck integration: 2 passed, including real source-chunk rendering, queue submission, and native asynchronous GPU picking.
• yarn website:build: passed.
• (cd website && yarn build): passed.
• All 46 example TypeScript checks and all seven bundle-size ceilings passed.
• GitHub Actions: bundle-size, all three browser coverage shards, examples, website, and aggregate tests passed on the final head.

Risks

• The compatibility bridge exists only for the current Deck 9.3 missing-language call; it restores global state immediately and should be removed after upstream Deck supplies its explicit shader language.
• This example is intentionally WebGPU-only and remains entirely outside production graph-module imports and dependencies.
• Graph readback occurs only for explicit deck.gl picking results; source columns and analytic/render buffers remain resident.