Add Swift port of V-JEPA2 with mlx-swift and camera apps

[gaarutyunov]

Major additions:

• Complete Swift port of V-JEPA2 model using mlx-swift
• macOS app with camera integration for real-time action recognition
• iPhone app with camera-based video classification
• Comprehensive test suite with cross-language Python-Swift comparison
• CI/CD integration for automated testing on macOS runners
• Something-Something-V2 labels (174 action classes)

Swift Implementation:

• VisionTransformer: ViT-Large/16 encoder with video support
• AttentivePooler: Classification head with cross-attention
• Core modules: Patch embedding, positional encodings, attention, MLP
• Support for both image and video inputs
• RoPE attention implementation
• Dynamic position encoding interpolation

Applications:

• macOS: Desktop app with split-view interface, real-time preview
• iOS: Mobile app with full-screen camera and overlay predictions
• Both apps include FPS monitoring and inference time metrics
• AVFoundation integration for camera capture

Testing:

• Unit tests for all components (shape verification)
• Cross-language tests comparing Python vs Swift outputs
• Component tests: embeddings, attention, MLP, blocks
• Integration tests: end-to-end classification pipeline
• Test data export from Python for cross-validation

CI/CD:

• GitHub Actions workflow with Swift test job
• Runs on macOS-14 with Apple Silicon
• Automated test data export and cross-language validation
• Swift package build and test automation

Performance:

• Native Apple Silicon optimization via MLX
• Real-time inference (~6.5 FPS on M2)
• Efficient memory management for video processing