feat(pai): add radar sensor support to PAI converter and ncore_vis

[janickm]

Summary

• Extend the PAI data converter to discover and convert up to 19 radar sensors per clip into NCore V4 RadarSensorComponent format
• Add radar visualization to ncore_vis (3D point cloud component with fusing, motion compensation, and radar-specific color styles including radial velocity, RCS, SNR)
• Add radar projection overlay to the camera component in ncore_vis
• Generalize ncore_project_pc_to_img.py to auto-detect and support radar sensors via RayBundleSensorProtocol

Data Signal Mapping

Per-detection radar signals from the PAI parquet are mapped as follows:

PAI Column	NCore Field	Notes
azimuth + elevation	direction (float32 [N,3])	Spherical → unit Cartesian
distance	distance_m (float32 [1,N])	Single return
radial_velocity	generic_data["radial_velocity"]	Doppler velocity (m/s)
rcs	generic_data["rcs"]	Radar Cross Section (dBsm)
snr	generic_data["snr"]	Signal-to-Noise Ratio (dB)
exist_probb	generic_data["exist_probb"]	Normalized uint8→float [0,1]
doppler_ambiguity, radar_model	generic_meta_data	Scan-level scalars

Key Implementation Details

• Radar extrinsics are only present in the non-offline sensor_extrinsics parquet (the .offline variant only contains cameras + lidar). The converter falls back to the raw variant automatically.
• Radar scans are grouped by scan_index (DataFrame index level 0), with each scan becoming one store_frame() call.
• The ncore_vis radar component auto-discovers generic_data fields and offers them as color styles, defaulting to radial_velocity when available.

Test

Verified end-to-end with streaming conversion of clip 0683f871-7d49-4bf9-8201-6227af1221bc:

Loaded metadata: 7 cameras, 1 lidars, 9 radars
Loading non-offline extrinsics for 9 radar sensors missing from offline calibration
Radar radar_front_center_srr_0: 12 frames converted
...
(9 radar sensors total, each producing ~12 frames per second of data)