feat(streampetr): project 2D annotations and fold augmentations into ego poses

[vividf]

Summary

Everything the auxiliary 2D head needs from the data side, plus augmentation-consistent ego poses for the temporal memory.

Changes

• LoadAnnotations2DFromBoxes3D (transforms/camera/annotations2d.py): projects the (already augmented) 3D gravity-center boxes onto every camera and emits per-camera gt_bboxes_2d / centers_2d / gt_labels_2d. Vectorized over boxes; corners behind the camera are masked out of the 2D extent; a box is kept only when its projected center lands inside the image (matching the reference recipe — a clamped center could fall outside its own clipped box and distort the center-based assignment); boxes are clipped to the image and zero-area slivers dropped. Runs after all geometric augmentations so the projection matrices and pixels agree.
• Ego-pose folding (transforms/geometry3d.py + camera / camera-lidar geometry): a scene augmentation A updates ego_pose' = ego_pose @ A⁻¹ (and the inverse accordingly), keeping lidar→global consistent — exactly what StreamPETR's memory warp (ego_pose_inv(t) @ ego_pose(t-1)) needs under train-time augmentation.
• Camera-order shuffling (transforms/camera/loading.py): per-frame train-time shuffle following the reference shuffle_cameras=True; images, intrinsics, and extrinsics are reordered together by construction.

Testing

Projective edge cases are unit-tested (box behind the camera, center leaving the image, partial visibility/clipping, multi-camera assignment, empty and malformed GT), plus ego-pose round-trip tests and a velocity-scaling regression test.

Stack 6/8 — base: feat/streampetr-training-infra.