Fix nearest-depth pixel sampling during TSDF integration

[kvnloo]

Fork-local clean upstream handoff for #7580.

This branch contains one commit on the exact current upstream main SHA b6c5e196384ad71e75b6e6f9c5da22d046221f1d.

It preserves the final file blobs validated in fork PR #2, while collapsing the experimental / formatting / CI-fix history into one reviewable commit.

Change

Round projected depth coordinates before the nearest-pixel lookup in TSDF integration.

Evidence inherited from the identical file blobs in #2

• focused pixel-selection regression: unchanged implementation 12 / 20, candidate 20 / 20;
• analytic fixed-ray surface regression: four tilted planes fail on the unchanged implementation and pass on the candidate; flat control remains effectively exact;
• broader 24-condition A/B: every tilted condition improved in RMS and absolute mean bias;
• corrected reconstruction-size fingerprints are deterministic on observed CPU runs;
• no bilinear sampling, renderer calibration, or public API changes.

Accelerator execution is not claimed where the fork workflow only exercises skip / CPU fallback paths.

CI on this PR is the gate for the squashed handoff itself.

AI-assisted implementation, experiment design, and validation review.

[update-docs]

Thanks for submitting this pull request! The maintainers of this repository would appreciate if you could update the CHANGELOG.md based on your changes.