Add Apple Silicon MPS inference backend

[natsusorahoshinochan-max]

Summary

This PR adds a pluggable inference backend layer and enables Apple Silicon GPU acceleration through PyTorch MPS while preserving existing CUDA/CPU behavior.

Changes

• Add backend abstraction for CPU, CUDA, MPS, and CoreML/ONNX providers
• Add Apple Silicon MPS paths for VR, MDX, MDXC, and Demucs v3/v4
• Keep CUDA as the first auto-selected backend when available
• Keep Demucs v1/v2 on the existing CPU/CUDA behavior
• Add smoke test harness for real-model inference validation
• Add macOS arm64 PyInstaller build support
• Update dependency rules so macOS does not install onnxruntime-gpu
• Ignore local build artifacts, venvs, model weights, and AppleDouble files

Validation

• CLI smoke summary: 27/27 passed on Mac mini m4
• Functional MDX23C test used Apple GPU (MPS) with fallback=false
• macOS arm64 .app build completed and passed ad-hoc codesign verification

[DOISolutions]

Cool