Add PnP and keypoint geometry across poses/points2D/pinhole

[oarriaga]

What

Consolidates the PnP and keypoint geometry that was scattered across the
applications layer and the pix2pose example into a single backend module
paz.keypoints, and ports the missing 2D keypoint-transform helpers from
master.

Before this, backend geometry (solve_PnP, solve_PnP_RANSAC,
project_points3D, build_cube_corners, Pose6D, the solver constants) lived
inside paz/applications/pose_estimators.py, and the pix2pose example carried
its own solve_pose_from_nocs with an inline Rodrigues conversion. There was no
paz/backend/keypoints.py at all, and the keypoint-transform layer that
master's keypoint pipelines are built from was absent.

Changes

• New paz/backend/keypoints.py (exposed as paz.keypoints):
  • PnP toolkit moved out of applications/pose_estimators.py: Pose6D, solver
    constants, solve_PnP, solve_PnP_RANSAC, project_points3D,
    build_cube_points3D.
  • Missing keypoint transforms ported from master as JAX (pure array math):
    normalize_keypoints2D, denormalize_keypoints2D, rotate_point2D,
    transform_keypoint, flip_keypoints_left_right, translate_keypoints,
    uv_to_vu. cv2 is reserved for the actual PnP solve only.
  • solve_pose_matrix_RANSAC: the reusable dense-correspondence → 6D-pose
    kernel lifted from the pix2pose example (RANSAC PnP, then Rodrigues to a
    rotation matrix). rotation_vector_to_matrix helper.
• applications/pose_estimators.py now imports the geometry from the
  backend; keeps only HeadPoseKeypointNet2D32, build_face_points3D, and
  draw_boxes3D.
• pix2pose example (pipeline.py, demo.py, validate.py) imports from
  paz.keypoints; solve_pose_from_nocs now just builds correspondences and
  calls solve_pose_matrix_RANSAC.
• Renamed the dimension-based cube builder to build_cube_points3D (master's
  name) to avoid colliding with the pre-existing
  pinhole.build_cube_corners, which takes min/max extents instead of
  width/height/depth.

Why

• Puts geometry in the matching backend per the agent guide ("new functions in
  the matching backend"), fixing the layering the pix2pose PR introduced.
• Delivers the keypoint-transform layer needed to port the keypoint_estimation
  training/augmentation pipeline next.
• Removes a confusing name collision (build_cube_corners meaning two different
  things).

Verification

• pytest paz/backend/keypoints_test.py — 10 pass (JAX transforms vs numpy
  reference, normalize/denormalize round trip, PnP + RANSAC pose recovery to
  1e-4, below-minimum returns None).
• pytest paz/losses/pix2pose_test.py paz/callbacks/evaluate_pose_test.py paz/backend/evaluation_test.py — 15 pass (no regression in the pose stack).
• pyflakes clean on all changed files; paz.applications and the head-pose
  app import unchanged; pinhole.build_cube_corners untouched.

[oarriaga]

Follow-up commit 098a3683: added the keypoint-aware augmentation training pipeline functions (master's RandomKeypointRotation / RandomKeypointTranslation ported as JAX) and adopted them in an example.

• paz.image.translate_image — pure-image (x, y) pixel shift via affine_transform, next to rotate.
• paz.keypoints.rotate_keypoints2D, rotate_image_and_keypoints, translate_image_and_keypoints, image_center2D — transform an image and its keypoints together so labels track the content.
• examples/probabilistic_keypoint_estimation/train.py — KeypointSequence now does rotate + translate + brightness through these backend helpers (was image-brightness only) and normalizes per batch.

Verified with hot-pixel tests (keypoint lands on the transformed pixel) and a 6-seed end-to-end check; pytest paz/backend/keypoints_test.py paz/backend/image_test.py → 22 pass.

[oarriaga]

Follow-up 4e13dcc2: reworked the layout after review. Instead of a new paz.keypoints backend, the geometry is distributed into the backends that already own each math domain (the earlier monolith duplicated points2D.transform, points2D.denormalize, and poses.project_to_image):

• paz.poses — Pose6D, PnP solver flags, solve_PnP, solve_PnP_RANSAC, rotation_vector_to_matrix, solve_pose_matrix_RANSAC, project_points3D.
• paz.points2D — normalize/denormalize_keypoints2D, rotate_point2D, rotate_keypoints2D, flip_keypoints_left_right, translate_keypoints, uv_to_vu (transform_keypoint dropped — the existing batched points2D.transform covers it).
• paz.pinhole — build_cube_points3D (beside build_cube_corners).

paz.keypoints is removed. Tests moved to poses_test, points2D_test, pinhole_test. pytest over the affected suites → 29 pass; the probabilistic-keypoints example still tracks keypoints through augmentation (6/6 seeds).