VBM select-based inverse-map decode

[swahtz]

Summary

This PR replaces the method in inverse-map decode in the NanoVDB VoxelBlockManager (tools/cuda/VoxelBlockManager.cuh) with an output-centric select. decodeInverseMaps previously had every thread in the block sweep all 512 voxel slots of every leaf overlapping the block — O(nLeaves × 512) work per block. It now runs one thread per output slot: each slot ranks itself into its leaf with a jumpMap popcount, then locates its voxel via the leaf's 9-bit mPrefixSum plus an in-word __fns bit select — O(1) per slot.

The public signature is unchanged and the decoded maps are byte-identical, so every downstream stencil consumer inherits the speedup for free. Benchmarked A/B on an RTX PRO 6000 (sm_120; 50 iters / 15 warmup).

What changed

The decode is, in effect, a rank + select over the VBM's bit-vectors:

• rank into the block's jumpMap (a popcount of leaf-start bits at or before the slot) selects which leaf owns the slot;
• select into that leaf's 512-bit value mask — guided by the precomputed 9-bit prefix-sum directory to find the containing 64-bit word, then an in-word __fns — recovers the voxel offset.

This removes the old sweep's data-dependent inner branch (if (index ∈ block)) and its per-leaf full scan, and makes the decode's work independent of leaf sparsity.

A/B — decode (dragon, median; BlockWidth = 2^log2width)

width	base (sweep)	this PR (select)	speedup
64	0.372 ms	0.171 ms	2.17×
128	0.288 ms	0.146 ms	1.97×
256	0.262 ms	0.162 ms	1.62×
512	0.281 ms	0.198 ms	1.42×

Decode is the shared front-end for all stencil consumers, so whole-consumer times improve too (dragon): 7-pt Laplacian −10…−15%, 27-tap box / 19-tap WENO −4…−9%, with no consumer regressing.

Validation

• Byte-identical decode maps: the VBM decode / box / lap / weno A/B comparisons all passed unchanged at widths 64/128/256/512 across dragon/emu/crawler/wdas_cloud.
• Final partial blocks and the beyond-last-active-voxel sentinels verified at all four widths.