feat: add MPS (Apple Silicon) support for inference

[ZimengXiong]

Support MPS backend for inference

• Updated device selection logic to prefer CUDA if available, then MPS, then CPU in both src/app.py and src/tool/edit_rgba_image.py. [1] [2]
• Set tensor data type to torch.bfloat16 for CUDA/MPS devices and torch.float32 for CPU. [1] [2]
• Updated model pipeline initialization and .to() calls to use the new device and data type logic in both files. [1] [2]
• Modified the generator in the infer function to use the selected device rather than hardcoding 'cuda'.