Fix: Graceful fallback for torch.jit.script on AMD/ROCm

[0xDELUXA]

@torch.jit.script fails on AMD/ROCm backend due to incomplete TorchScript support. This replaces the decorators on fast_diag_generalized_box_iou and fast_diag_box_iou with a try/except block that attempts JIT compilation at import time and silently falls back to eager mode if it fails.

Behavior on CUDA and CPU is unchanged.

Without this change, AMD/ROCm users encounter the following error:

Error loading C:\ComfyUI\custom_nodes\ComfyUI-RMBG\py\AILab_SAM3Segment.py: module 'torch.distributed' has no attribute 'rpc'

However, torch.distributed.rpc is actually available:

(venv) PS C:\ComfyUI> python -c "import torch; print(hasattr(torch.distributed, 'rpc'))"
True