Add MMPose wholebody estimator

[bricksdont]

Summary

This PR adds support for running MMPose wholebody inference on a video and converting the result into the library's .pose format.

• pose_format/utils/mmposewholebody.py — new module with estimate_mmpose_wholebody(input_path, fps, width, height). Runs MMPoseInferencer('wholebody') on a video file and returns a Pose object using the canonical COCO Wholebody 133 header (from cocowholebody133_header.py, introduced in Add COCO Wholebody 133 keypoint schema and generic.py support #226). Frames with no detected person are represented as fully-masked zero rows, preserving temporal alignment with the source video.
• pose_format/utils/mmposewholebody_test.py — 8 unit tests covering header structure, output shape, masked empty frames, and the all-empty-video case. MMPose is stubbed via sys.modules so tests run without an actual MMPose installation.
• pyproject.toml — new mmpose optional install target (pip install pose-format[mmpose]) with version floors derived from a known-good OpenMMLab 2.x combination.

Design notes

Not wired into the CLI — video_to_pose / videos_to_poses currently only support mediapipe. MMPose could be added to the CLI later; this PR intentionally keeps that scope separate.

Format vs. estimator — MMPose wholebody outputs COCO Wholebody 133 keypoints. A .pose file produced by this function is structurally identical to one from any other COCO Wholebody 133 estimator (OpenPifPaf, SDPose, etc.). detect_known_pose_format correctly identifies such files as "coco_wholebody_133". MMPose is therefore an estimator that produces an existing named format, not a new format in its own right.

mmcv GPU install caveat — plain pip install mmcv installs a CPU-only build; GPU users need the OpenMMLab CUDA-specific index. The ImportError message in the module links to the mmcv install guide.

Relationship to other PRs

This builds on #226 (COCO Wholebody 133 canonical header). Future PRs will add OpenPifPaf and SDPose estimators using the same header.

Test plan

• pytest src/python/pose_format/utils/mmposewholebody_test.py -v passes without MMPose installed
• pytest src/python/pose_format/utils/generic_test.py -v passes (no regressions)
• With MMPose installed: estimate_mmpose_wholebody("video.mp4", fps=25, width=1920, height=1080) returns a Pose with detect_known_pose_format → "coco_wholebody_133"

🤖 Generated with Claude Code

[bricksdont]

mmpose dependencies are rather heavy, so they are excluded from the tests / CI