Reads 32-bit float scene slices from *-TERRAIN-DISP-MERGED.tiff and
per-scene poses from *-INTERFRAME.corr-xml, outputs COLMAP text-format
sparse model (cameras.txt, images.txt, empty points3D.txt) plus
16-bit PNG images for OpenMVS ingestion.

Camera: PINHOLE, fl derived from 32° HFoV (1116 px for 640 px width).
Rotation: mirrors GeometryCorrection.getCommonRotMatrix(); convention
verified in Phase 3 Blender smoke test.
Co-authored-by: Claude <claude@elphel.com>