Add optional CUDA LiDAR terrain sweeps

[jpooley-hearst]

Summary

• add an optional standalone CUDA terrain-sweep executable without changing the existing CPU build
• prepare GDAL-readable LiDAR DTM/DSM inputs as bounded local AEQD rasters with explicit nodata and provenance
• cache sector-major terrain profiles once and reuse them across as many as 16 explicitly configured pattern/power/height scenarios
• support RAM-bounded disk spooling, exact 1 kW base-field export, azimuth shards, GeoTIFF conversion, and census aggregation
• add neutral examples, validation documentation, Python regression coverage, and CI for the terrain/output tools

Design

The contribution lives under cuda/ and builds as signalserver-cuda. The existing src/Makefile, signalserver, signalserverHD, and signalserverLIDAR paths are unchanged.

A run shares transmitter location, frequency, terrain, radius, and sampling geometry. Terrain profiles and prefix state are reused; each [variant.NAME] is an explicitly requested scenario, so pattern and power inputs are not cross-joined.

The LiDAR preparation path accepts GeoTIFF, COG, VRT, and other GDAL-readable sources. It creates signed-int16 DTM and optional uint8 DSM-minus-DTM clutter rasters, validates metadata and binary sizes, and carries nodata through cache, GPU propagation, and output.

Validation boundary

Free-space output has an analytic regression baseline. The included CUDA P.1812 and ITM implementations remain experimental and emit model_validation=unvalidated_implementation plus a runtime warning. The documentation explicitly excludes them from regulatory, safety, or other decision-grade use until independent reference-vector validation is complete.

Verification

• CUDA 13.2 / MSVC Release configure and build
• 16 Python/GDAL terrain, GeoTIFF, merge, and census regression tests
• synthetic GPU suite covering cache build/reuse, DTM/CHM nodata, direct versus spooled output, exact base-field export, complementary azimuth shards, long profiles, edge masking, and experimental-model warning metadata
• final smoke invariant: direct=spool, shards=full
• staged diff audit for private names, local paths, secrets, generated binaries, bytecode, empty files, and whitespace errors