Improve handling of multi-dimensional features in Injection operations

[swahtz]

📋 Originally created by @sifakis

Original author: @sifakis
Date: 2025-07-14T23:09:05Z

---

The current handling of injection for sidecars of multidimensional and/or strided features (https://github.com/NVIDIA-Omniverse/openvdb/pull/636) performs an expensive and maybe unnecessary translation from a flattened linear offset of a scalar entry within a tensor (as if it was contiguous and row-major) into the corresponding offset under a specific stride pattern. It should be possible to circumvent that computation (involving an integer division for every single scalar copy), and perform this computation purely using the voxel index offset.

[fwilliams]

@sifakis I don't remember if we go to this before the last release