OOM error (yes, another one)

### 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

No out of memory error

### Actual Behavior

Out of memory

### Steps to Reproduce

Works with any version of ComfyUI to date and any gpu memory amount. The amount just determines how fast you get this OOM.

Assumption: 32GB VRAM.

1. Take the default Krea workflow.
2. In the latent image set the **batch size** to 20.
3. It works all fine.

Now repeat but **change the batch size** to 30. OOM. Perhaps as expected.
Now set the batch size back to 20. Remember, this worked! You get an OOM. This should not happen!

### Debug Logs

```powershell
Irrelevant, it's reproducible with the above under any config.
```

### Other

My assumption here is that setting a batch size somewhere reserves memory that isn't cleaned when you change the batch size. Like when i'd set a batch size of 30 first (which gives the OOM), setting it at 20 again still has the memory allocation of the batch size of 30. Just a guess, i could be wrong.

And yes, this is on a clean comfy with no nodes installed and just the models installed that the workflow needs.

Start by reproducing the issue with the default Krea workflow: run batch size 20, then 30 to trigger OOM, then return to 20 without restarting. Investigate how GPU memory is handled when the batch size changes; done means the batch-20 workflow runs again after the failed batch-30 attempt without requiring a restart.

Written by the indexing model from the issue text.