The data handler and saved outputs

FreemocapDataHandler wraps the loaded arrays (body, right hand, left hand, face, plus other components like center of mass) with name-indexed accessors, shape validation on every setter, processing-stage snapshots (mark_processing_stage keeps deep copies you can retrieve later), whole-dataset rotate/translate/scale transforms, and extract_data_from_empties(), which reads trajectories back out of the Blender scene frame by frame. Two known quirks in the current code: applying scale() to the whole dataset (the default, with no component name) reaches an other-component branch in FreemocapDataTransformer that multiplies the component object itself instead of its .data array and raises, and the other branch of extract_data_from_empties() constructs an np.ndarray from a Blender location object instead of using np.array, which raises and aborts the whole extraction whenever an other group is present in the empties mapping (body, hand, and face components extract correctly).

FreemocapDataSaver writes a saved_data/ folder into the recording: _FREEMOCAP_DATA_README.md documenting the layout, info/freemocap_data_handler.pkl (a pickle of the whole handler), info/metadata.json, info/trajectory_names.json, an npy/ folder of (frame, name, xyz) arrays including a concatenated all_frame_name_xyz.npy, a csv/ folder with per-component trajectories, and a combined all_trajectories.csv written directly into saved_data/ rather than inside the csv/ folder. CSVs are written with np.savetxt specifically so this works inside Blender without pandas.