Add segmented object affine transformation module with benchmarking

[rojas70]

Motivation

• Provide a segmentation-first pipeline to apply Euclidean affine transforms to target objects in images rather than warping entire frames.
• Couple image-space manipulation with robot planar state updates so pixel-space transforms can inform external robot state.
• Offer a simple performance study to decide between full-frame vs ROI-only warping strategies.

Description

• Add franka_env.utils.image_object_affine implementing EuclideanAffine2D, build_affine_matrix, preprocess_mask, transform_segmented_object, update_planar_robot_state, and benchmark_affine_strategies using NumPy and scipy.ndimage for warping and morphology.
• Implement inverse-mapped warping, nearest-neighbor/linear sampling via ndimage.map_coordinates, background hole-filling via distance_transform_edt, and mask denoising with morphological close/open.
• Export the new utilities from franka_env.utils by updating __init__.py.
• Add unit tests in serl_robot_infra/tests/test_image_object_affine.py covering object movement, robot-state coupling, and benchmark metrics.

Testing

• Ran pytest -q serl_robot_infra/tests/test_image_object_affine.py, which reports 3 passed after replacing an initial OpenCV-based implementation with a SciPy-based one to avoid libGL.so.1 runtime issues.
• Benchmarks API benchmark_affine_strategies is exercised by the tests and returns the expected keys (full_frame_ms, roi_only_ms, speedup_x).

---

Codex Task