feat(nuscenes/lidar-model): add nuScenes V4 converter and lidar-model estimation lib / evaluation tool

[janickm]

Summary

Add a nuScenes dataset converter for NCore V4 format, a generic structured lidar model extraction library, and a lidar model evaluation tool.

Closes #123

Architecture

The work is split into three self-contained commits:

1. feat(data_converter): add structured lidar model extraction library

tools/data_converter/structured_lidar_model.py -- a generic, reusable library for deriving structured spinning lidar models from motion-compensated point cloud data. Designed for datasets that lack raw sensor timestamps.

Key features:

• Composable steps: extract_column_azimuths, compute_column_alignment, assign_model_columns, compute_frame_timestamps, decompensate_frame
• Resolution upsampling (upsample_model): interpolate column azimuths to 2x/4x for sub-column alignment precision
• Multi-frame optimization (optimize_model): median-based correction across many frames
• Pre-computed timestamps support: align_frame(timestamps_us=...) for sensors with known per-point timestamps
• HDL-32E presets: spec elevations, analytical firing offsets, nominal model factory
• 18 unit tests (no external data dependency)

2. feat(tools): add lidar model evaluation tool

tools/ncore_evaluate_lidar_model.py -- CLI tool that evaluates model quality by comparing model-predicted vs stored native directions. Works on any V4 sequence.

Outputs:

• Numerical metrics: angular error (mean/median/p95), per-row and per-frame breakdowns, pixel-equivalent error
• Overlay images: native (cyan) + model (red) points on camera
• Error heatmap images: points colored by angular error magnitude

3. feat(nuscenes): add nuScenes to NCore V4 converter

Full converter handling 6 cameras, 1 lidar (HDL-32E), 5 radars, 3D cuboid annotations. Uses the generic library directly (no local lidar model code).

Lidar Model Quality

Dataset	Mean error (far)	Pixel equiv	Notes
Waymo (ground truth)	0.010 deg	0.35 px	Theoretical floor
PAI (extraction)	0.028 deg	~0.5 px	Factory-quality
nuScenes (4x nominal + opt)	0.029 deg	0.7 px	This PR
nuScenes (1x empirical + opt)	0.083 deg	1.8 px	Without resolution upsampling

Recommended config: --lidar-model-source nominal --lidar-model-resolution 4 --lidar-model-optimization-passes 1

Key Design Decisions

1. Resolution upsampling is the dominant improvement (0.10 -> 0.03 deg). Column alignment quantization was the bottleneck.
2. Nominal model from spec avoids circular dependency (need timestamps to decompensate, need model to derive timestamps).
3. Analytical firing offsets (from HDL-32E timing) replace unreliable empirical measurements for ground-hitting beams.
4. timestamps_us parameter on align_frame supports sensors with known per-point timestamps (e.g., raw PCAP data).
5. Library lives at tools/data_converter/ level (not inside nuscenes/) for reuse by other converters.

Testing

# Unit tests (no external data)
bazel test //tools/data_converter:pytest_structured_lidar_model_3_11

# Integration test (requires nuScenes mini)
bazel test //tools/data_converter/nuscenes:pytest_converter_3_11 \
    --test_env=NUSCENES_DIR=/path/to/nuscenes/mini/v1.0-mini \
    --test_env=NUSCENES_VERSION=v1.0-mini

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[janickm]

/ok to test d123655