Replace dlib with MediaPipe for face detection and add ONNX export

[AmirhosseinSoltan]

What changed?

Great inference speed by:

• Replaced dlib CNN face detector with MediaPipe FaceDetection
• Added one-time ONNX export pipeline: PyTorch weights
• Switched inference from PyTorch forward pass to ONNX Runtime CPUExecutionProvider
• Added --onnx CLI argument to override the ONNX model path

Why?

The original pipeline was far too slow for real-time CPU inference. The dlib CNN face detector alone took 100–300 ms/frame — 3–10× the entire frame budget for 30 FPS — making the bottleneck the detector, not the ResNet. Replacing it with MediaPipe brings detection down to ~5–10 ms/frame. Exporting to ONNX reduces ResNet-50 inference time by ~2× and model size from 91 MB to ~23 MB, with negligible accuracy impact on a binary eye-contact score.