fix(radix-sort): OneSweep on NVIDIA, indirect bind fix, GSplat bit alignment

[mvaligursky]

Summary

Improves automatic OneSweep selection on NVIDIA WebGPU, fixes indirect OneSweep bind group creation, aligns GSplat GPU radix sort key bit counts with the active sorter radix width, and updates the indirect radix sort example so it matches production prepareIndirect / multi-slot behaviour.

Changes

• compute-radix-sort.js: _canUseOneSweep no longer rejects NVIDIA when the adapter reports a high maxSubgroupSize (D3D12 ceiling) or omits subgroup size (minSubgroupSize 0). Only refuses when minSubgroupSize > 32, which would break 32-bit subgroup ballot masks in the binning shader. Clarified Debug.warnOnce for explicit RADIX_SORT_ONESWEEP on non-validated hardware.
• compute-radix-sort-onesweep.js: Constructor assert aligned with the same minSubgroupSize policy. Indirect bind groups: use three separate BindStorageBufferFormat('b_sortElementCount') instances so BindGroupFormat slot assignment is not overwritten across histogram / scan / binning layouts (fixes WebGPU binding index 7 not present validation errors).
• gsplat-manager.js: In sortGpu, round numBits with gpuSorter.radixBits instead of hardcoded 4 so OneSweep receives a multiple of 8. JSDoc for dispatchGpuSort updated accordingly.
• radix-sort-indirect-compute.example.mjs: ComputeRadixSort constructed with { indirect: true }; prepare pass uses prepareIndirect() metadata, reserves slotCount dispatch slots, and writes per-slot workgroup counts with correct granularities; UI bit options filtered by radixSort.radixBits.

Testing

• ESLint on the four changed files.
• WebGPU NVIDIA: run radix-sort-indirect-compute example — no bind validation errors, no device hang.
• GSplat GPU sort path: sort key bit count valid for both multipass (4) and OneSweep (8).

Public API

None.