Add a COLMAP export format to polyform (closes #9)

[kunjikakgp]

Closes #9 ("How to get the raw data into Colmap format to use 3DGaussian?"). Also relevant to #11 (LiDAR Point Cloud Alignment) and #13 (Alignment between camera poses and generated mesh) -- same root need, though this PR focuses specifically on #9.

Summary

Adds a COLMAPConvertor (alongside the existing InstantNGPConvertor) that exports a Polycam capture as a standard COLMAP text model (cameras.txt, images.txt, points3D.txt). This is the format expected by the reference 3D Gaussian Splatting implementation and by Nerfstudio's COLMAP data parser, enabling direct use with COLMAP-based tooling without re-running COLMAP's feature matching and SfM pipeline, since Polycam already provides globally optimized camera poses..

Usage, matching the existing CLI pattern:

python3 -m polyform.convert --format colmap

or, for a single shared camera instead of one-per-image:

python3 -m polyform.convert --format colmap --shared_camera

What's included

• polyform/convertors/colmap.py -- the convertor itself, plus a standalone arkit_pose_to_colmap() helper that converts Polycam's raw ARKit camera-to-world transform into COLMAP's world-to-camera quaternion + translation.
• polyform/convert.py -- wires --format colmap into the existing CLI entry point.
• README.md -- new "COLMAP" section (mirrors the existing "Instant-ngp" section's structure and tone).
• tests/ (new) -- 18 pytest tests covering the coordinate-conversion math and the end-to-end convertor, using synthetic capture folders since the repo doesn't ship any sample data.

Design decisions

Axis-convention conversion: Polycam's raw transform_rows (rotate=False) is a camera-to-world matrix in ARKit/OpenGL camera axes (+X right, +Y up, +Z out of the screen). COLMAP/OpenCV expect +X right, +Y down, +Z into the scene, stored as world-to-camera rather than camera-to-world. The conversion flips the local Y/Z axes of the camera-to-world rotation, then transposes to get world-to-camera. I deliberately did not reuse the existing rotate=True branch in Camera.init (the one InstantNGPConvertor uses) -- that's a different axis permutation tuned for instant-ngp/nerfstudio's convention, not COLMAP's, and mixing the two would silently produce wrong poses.

Rotation matrix to quaternion: implemented by hand (Shepperd's method) rather than adding scipy as a new dependency, since polyform currently only depends on numpy, fire and Pillow. A naive sqrt(trace)/2 formula is numerically unstable close to 180-degree rotations, which do happen in real scan trajectories (e.g. looping around an object).

Per-image cameras by default: InstantNGPConvertor already notes that Polycam's per-frame intrinsics vary by a pixel or two frame to frame. I followed that precedent -- every image gets its own COLMAP camera by default, with an opt-in shared_camera=True (--shared_camera on the CLI) for workflows that assume one shared camera.

points3D.txt is intentionally empty: Polycam's raw export doesn't include COLMAP-style per-image 2D/3D correspondences (that's normally the output of the feature-matching + triangulation step this convertor lets people skip). An empty points3D.txt is valid input for tools that only need camera poses -- 3D Gaussian Splatting falls back to random point initialization when it's empty. Documented explicitly in the README.

Possible follow-ups (not in this PR)

• Seed points3D.txt from the depth maps Polycam already exports (keyframes/depth/*.png), by back-projecting a subsample of pixels into a sparse point cloud. Didn't attempt this here since I don't have a real Polycam capture to confirm the exact depth encoding (units, 16-bit scale factor) against, and didn't want to ship unverified back-projection code.
• Binary model export (colmap model_converter can already do this conversion, so kept out of scope here).

Testing

Ran via PYTHONPATH=. python3 -m pytest tests/ -v -- 18 passed. Also ran flake8 (--extend-ignore=F403,F405,E305 to match the star-import and spacing conventions already used in instant_ngp.py) with no findings on the new files.

Coverage: rotation-matrix/quaternion round-tripping (including a rejection check for non-rotation-matrix input), the ARKit->COLMAP axis conversion in isolation, and end-to-end convertor behavior (correct files written, per-image vs shared_camera, unit-norm quaternions, empty points3D.txt, optimized-vs-raw pose handling with cropped intrinsics, empty-folder no-crash, default output path). Also manually verified end-to-end via the CLI against a small synthetic capture.