perf(data-gpu): fix O(all) voxel scan in physicsRenderBridge, add static-instance fast path

[krisnye]

Summary

• physicsRenderBridge excluded voxel-tagged and already-resolved rows with a per-row store.read/store.get check inside the scan loop, so it visited every row of every matching archetype each frame — including large static-collider populations (e.g. a terrain heightfield) that were always skipped anyway. The exclusion now happens at the archetype-query level (exclude: ["voxelShapeName", "mesh"]), so those rows never enter the loop (O(pending) instead of O(all)).
• pbrFactorRender gains an opt-in _pbrFactorStaticInstances resource: once a scene's drawable count stabilizes, cache per-mesh instance buffers/bind groups and just re-issue drawIndexed on later frames, skipping the per-frame TRS gather + instance-buffer upload for scenes that never move or grow (e.g. baked terrain).
• Bump all publishable package versions to 0.10.21 in lockstep (pnpm bump).

Found while profiling a 256×256 voxel terrain sample in a downstream app: physicsRenderBridge was costing 13-15ms/frame (the majority of CPU frame time) purely from scanning ~65k static terrain colliders it was always going to skip.

Test plan

• pnpm --filter @adobe/data-gpu typecheck — passes (pre-existing, unrelated interpolate.ts error confirmed present on main before this branch too)
• Verified in a downstream app (mechs terrain-world sample): frame time dropped from ~33ms to ~17ms after applying the equivalent fix locally