BiRefNet 512x512, patched to run on ONNX Runtime WebGPU

Identical outputs to onnx-community/BiRefNet_512x512-ONNX (verified bit-exact). Five ScatterND nodes in /bb/layers.3_1/ received zero-sized updates tensors (the 12x12 map divides exactly by the swin window at 512 input), which ORT-web's WebGPU backend refuses to run. Since a scatter of nothing is a no-op with static shapes, those nodes are replaced with Identity and 530 dead nodes pruned. fp16 only.