SeedVR2 image upscaling breaks batched images

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

All images in an image batch are upscaled correctly

### Actual Behavior

The first image is fine, but subsequent images in the batch are visually broken

### Steps to Reproduce

1. Download the SeedVR2 image upscaling workflow (utility_seedvr2_7b_int8_upscale_image) from the official templates
2. Instead of feeding a single image, create an image batch (for example, using Image Batch Multi node in KJNodes) with 2 images and connect it to the image input of Upscale Image subgraph
3. Run the workflow

### Debug Logs

```powershell
[INFO] Requested to load SeedVR2
[INFO] Model SeedVR2 prepared for dynamic VRAM loading. 7944MB Staged. 0 patches attached. Force pre-loaded 576 weights: 3853 KB.
100%|███████████████████████████████████████| 1/1 [00:03<00:00, 3.41s/it, Model Initialization complete! ]
[INFO] Prompt executed in 16.55 seconds
```

### Other
(Using non-tiled VAE Encode/Decode didn't change the result)

Start by running the official utility_seedvr2_7b_int8_upscale_image workflow with a single image, then repeat with a two-image batch connected to the Upscale Image subgraph. Compare the outputs and investigate the SeedVR2 upscale path; done means every image in the batch is visually correct.

Written by the indexing model from the issue text.