Add cross-library TSDF / ESDF / Occupancy benchmark drivers

[fwilliams]

Summary

Adds a tests/benchmarks/tsdf_fusion/ benchmark suite that compares fvdb against the public sparse-voxel libraries (nvblox, VDBFusion, Open3D) on real and synthetic RGB-D / LiDAR workloads.

Pairs with the fvdb-core TSDF / ESDF / Occupancy / Decay PR (openvdb/fvdb-core#656, itself stacked on #655). The drivers exercise the new Grid.integrate_tsdf_frames, integrate_tsdf_from_points, integrate_occupancy_from_points, compute_esdf, and decay_and_prune APIs.

Per-dataset bench drivers (tests/benchmarks/tsdf_fusion/cross_library/)

Driver	Dataset	Workload
bench_open3d_vs_fvdb.py	Synthetic sphere	TSDF, fvdb vs Open3D
bench_replica_full.py	Replica	TSDF + ESDF, multi-scene, full frame rate
bench_esdf_replica.py	Replica	ESDF scale sweep
bench_kitti.py	KITTI Odometry	LiDAR TSDF
bench_esdf_kitti.py	KITTI Odometry	LiDAR ESDF
bench_occupancy_kitti.py	KITTI Odometry	LiDAR occupancy
bench_decay_kitti.py	KITTI Odometry	Dynamic-scene decay-and-prune
bench_mai_city.py	Mai City (LiDAR)	LiDAR TSDF, fvdb vs VDBFusion vs nvblox
bench_seven_scenes.py	7-Scenes (RGB-D)	TSDF, long-trajectory
bench_esdf_vs_nvblox.py	Mai City	ESDF, fvdb vs nvblox
bench_occupancy_vs_nvblox.py	Mai City	Occupancy, fvdb vs nvblox

Supporting modules

• kitti_loader.py, replica_loader.py, mai_city_loader.py, seven_scenes_loader.py — minimal per-dataset loaders that yield (intrinsics, cam_to_world, depth_or_points) tuples.
• download_kitti.py, download_replica.py, download_replica_zip.py — resumable, parallel-stream downloaders for the public dataset archives.
• nvblox_runner.py — Python wrapper around the nvblox CLI that matches the cross-library interface used by the bench drivers above.
• install_nvblox.sh, install_vdbfusion.sh — reproducible install recipes for the two C++ comparison libraries (each builds against the fvdb conda env's TBB / Blosc / Boost).
• data/README.md, data/.gitignore — conventions for placing dataset trees under data/. The dataset files themselves are not vendored.

Each driver accepts --help for its full argument surface and writes a JSON results file alongside per-frame stdout summaries.

Test plan

A no-external-dataset smoke test runs in a few seconds and is the easiest way to verify the setup:

python tests/benchmarks/tsdf_fusion/cross_library/bench_open3d_vs_fvdb.py \
    --n-frames 16 --json-out /tmp/smoke.json

The dataset-bound drivers expect the public archives under tests/benchmarks/tsdf_fusion/data/<dataset>/ (gitignored; downloaders provided in the same directory).

• Synthetic sphere smoke test runs end-to-end with both fvdb and Open3D code paths and writes the JSON results file.
• At least one dataset-bound driver runs end-to-end on its public archive (e.g. bench_replica_full.py against a Replica scene).
• install_nvblox.sh / install_vdbfusion.sh produce working installs from a clean conda env (these are scripts, not part of the fvdb wheel).