feat: add pyproject.toml (uv) and Apple Silicon MPS support

[Micuks]

Background

I was trying to set up Edit Banana on my MacBook (Apple Silicon) and ran into a few issues:

1. No pyproject.toml — had to manually pip install -r requirements.txt, and optional deps like torch/sam3 weren't structured, making it hard to manage with modern tools like uv.
2. SAM3 device detection only checks for CUDA, so it falls back to CPU on macOS even though MPS (Metal Performance Shaders) is available and much faster.

Changes

1. Add pyproject.toml

• Migrate deps from requirements.txt into a standard pyproject.toml
• Add optional dependency groups so users can install what they need:
  • pip install -e ".[torch]" — PyTorch + torchvision
  • pip install -e ".[sam3]" — SAM3 + torch + missing upstream deps (einops, pycocotools, psutil)
  • pip install -e ".[paddleocr]" / .[formula] / .[rmbg] — same as before
• Configure [tool.uv.sources] for local sam3_src/ editable install
• Update .gitignore to exclude uv.lock and *.egg-info/

2. MPS device support

• SAM3Model.__init__: auto-detect MPS when no device is specified (cuda > mps > cpu)
• config.yaml.example: document the "mps" device option and PYTORCH_ENABLE_MPS_FALLBACK=1 env var

Note: Running on MPS also requires patches to the upstream SAM3 source (bfloat16 → float32 cast, fused op fallback, pin_memory guard, etc.). These are tracked in facebookresearch/sam3#173 and need to be applied to the local sam3_src/ clone. I can add a patch script or document the steps if that would be helpful.

Test

Tested on MacBook Pro (M-series) with Python 3.10, PyTorch 2.11, uv 0.6:

$ PYTORCH_ENABLE_MPS_FALLBACK=1 uv run python main.py -i input/test.png
...
[SAM3Model] 加载模型中... (设备: mps)
[SAM3Model] 模型加载完成！
...
Done: 1/1 succeeded

Compatibility

• requirements.txt is kept as-is for users who don't use pyproject.toml
• CUDA users are not affected — CUDA is still the first choice when available
• The existing scripts/setup_sam3.sh workflow is unchanged