Does not work as intended (Small speedup for text model sampling.#14773 )

### Custom Node Testing

- [ ] 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
Prompt expansion should be consistent for identical workflow, model, seed, and prompt.

### Actual Behavior
After commit #14773, clip.generate() produces different prompt expansions, resulting in noticeably different image semantics. Reverting 000c6b78 restores the previous behavior.

Before commit revert.

After commit revert

### Steps to Reproduce

Use Gemma (gemma4_e4b_it_fp8_scaled) to expand / enhance a prompt.
Either use ComfyUI nodes or my attached one, behavior is the same

[gemma_T_Enh_CLIP_Presets.py](https://github.com/user-attachments/files/30392627/gemma_T_Enh_CLIP_Presets.py)

### Debug Logs

```powershell
nothing to report
```

### Other

_No response_

Compare commit #14773 and revert 000c6b78, focusing on changes affecting clip.generate(). Reproduce the issue with Gemma using the attached gemma_T_Enh_CLIP_Presets.py or equivalent ComfyUI nodes, keeping the workflow, model, seed, and prompt identical. Done means prompt expansion is consistent and matches the behavior before the regression.

Written by the indexing model from the issue text.