LTX2.5 is extremely slow during the decoding stage for 10-second videos.

### 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

I generate 1024×1024 5-second videos using the image-to-video mode, and they can be completed in around 60 seconds, with normal decoding times. However, once I extend the duration to 10 seconds, the total processing time increases to around 900 seconds.

After observing the process, I found that the initial processing stages finish very quickly, but the process gets stuck during the decoding stage, where it wastes several hundred seconds. Even after disabling the prompt enhancement feature, the issue persists.

My system has 24 GB of VRAM and 96 GB of RAM, with plenty of available disk space. The storage is an NVMe SSD with read speeds of around 7000 MB/s.

Please fix this extremely slow decoding bug as soon as possible.

### Actual Behavior

generate

### Steps to Reproduce

no

### Debug Logs

```powershell
[INFO] got prompt
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[INFO] Requested to load CausalDiffusionVAE
[INFO] loaded completely; 1403.92 MB loaded, full load: True
[INFO] Found quantization metadata version 1
[INFO] Using MixedPrecisionOps for text encoder
[INFO] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
[INFO] Requested to load LTXAVTEModel_
[INFO] Model LTXAVTEModel_ prepared for dynamic VRAM loading. 14612MB Staged. 0 patches attached. Force pre-loaded 293 weights: 1535 KB.
FETCH ComfyRegistry Data: 10/171
[INFO] Model LTXAVTEModel_ prepared for dynamic VRAM loading. 14612MB Staged. 0 patches attached. Force pre-loaded 293 weights: 1535 KB.
[INFO] Found quantization metadata version 1
[INFO] Detected mixed precision quantization
[INFO] Using mixed precision operations
[INFO] Native ops: float8_e4m3fn, int8_tensorwise, convrot_w4a4, asym_w4a8_int8, float8_e5m2 , emulated ops: nvfp4, mxfp8
[INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16
[INFO] model_type FLUX
[INFO] Requested to load LTXAV
[INFO] Model LTXAV prepared for dynamic VRAM loading. 20484MB Staged. 0 patches attached. Force pre-loaded 608 weights: 3303 KB.
0%| | 0/8 [00:00

No source files or tests are named. First reproduce the 5-second and 10-second image-to-video cases with custom nodes disabled, then compare the decoding timing with the VAE-related log entries; done means the 10-second video no longer spends several hundred seconds in decoding.

Written by the indexing model from the issue text.