[Task] keypoint-detection model support

Overview

Keypoint detection (a.k.a. human pose estimation) models predict the locations of anatomical landmarks — joints, facial landmarks, hand keypoints — on people in images or video frames. This enables downstream reasoning about body posture, gesture, and movement.

Agent Scenarios

• Fitness / sports coaching agent: analyze exercise form in real time, flag incorrect posture or track rep counts from a webcam feed
• Accessibility agent: interpret sign language or body gestures as input to a conversational agent
• AR / avatar agent: drive avatar animations by mapping detected pose to a virtual skeleton
• Safety monitoring agent: detect falls or abnormal postures in elderly care or industrial environments

ModelKit Integration

Models must pass the full wmk pipeline on all EPs:

wmk config → wmk build (ONNX export) → wmk perf → wmk eval

Acceptance Criteria

• usyd-community/vitpose-plus-base
• usyd-community/vitpose-base-simple
• usyd-community/vitpose-plus-small
• usyd-community/vitpose-plus-large
• usyd-community/vitpose-plus-huge
• stanfordmimi/synthpose-vitpose-huge-hf

Start with the wmk config, build, perf, and eval pipeline described in the issue, then inspect how existing models are integrated. Add support for the six listed keypoint-detection models and verify that each passes the full pipeline, including ONNX export.