Add FMPose3D integration for monocular 3D pose estimation

Add functionality for using https://github.com/AdaptiveMotorControlLab/FMPose3D to lift 2D predictions to 3D keypoints. (see paper)

Summary

• Integrate FMPose3D into the DeepLabCut model zoo, enabling monocular 3D pose estimation via flow matching for both human (17 H36M joints) and animal (26 Animal3D joints) pipelines
• Add a convenience wrapper get_fmpose3d_inference_api() that creates a ready-to-use inference object with automatic weight downloading from HuggingFace Hub
• Add integration tests covering API initialization (lightweight) and full end-to-end 2D-to-3D inference (network-dependent)
• Add a Jupyter notebook recipe to the docs cookbook showing end-to-end usage, two-stage workflows, and how to lift existing DLC 2D predictions to 3D

Details

Core integration (fmpose3d.py): A thin wrapper that constructs an FMPose3DConfig and returns an FMPose3DInference object. Supports both fmpose3d_humans and fmpose3d_animals model types, optional custom snapshot paths, device selection, and extra config kwargs.

Tests (test_fmpose_integration.py):

• test_api_init — parametrized over both model types; verifies the API object has the expected interface (prepare_2d, pose_3d, predict). Runs on CPU with no network required.
• test_prepare_2d_and_pose_3d / test_predict_end_to_end — full integration tests that download weights and run inference on the included example image. Skipped automatically when no network is available.

Recipe notebook (fmpose3d.ipynb): Covers three workflows:

1. End-to-end human pose estimation (single predict call)
2. End-to-end animal pose estimation
3. Two-stage pipeline (separate 2D detection + 3D lifting)
4. Lifting existing DLC 2D predictions (from analyze_images or analyze_videos) to 3D