- Add fmpose3d_humans and fmpose3d_animals framework mapping to pytorch.
- Branch early in video_inference_superanimal(...) when model_name.startswith("fmpose3d").
- Route directly to _video_inference_fmpose3d(...)

Add _video_inference_fmpose3d(...) as a dedicated path.

Keep batch loop minimal in DLC:
- result_2d = api.prepare_2d(np.stack(frames))
- result_3d = api.pose_3d(result_2d.keypoints, result_2d.image_size)

DLC responsibilities only:
- iterate video frames/chunks,
- convert 2D results to DLC {"bodyparts": ...} layout,
- save .h5, .json, _3d.json, and optional labeled video.