[ComfyUI v0.23.0] Sampling hangs at 0/N, GPU pinned 100%, with PromptRelay (LTX-2.3, Windows) — works fine on v0.22.0

### 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

## Summary
A custom node (kijai/ComfyUI-PromptRelay) injects an additive attention bias into LTX-2.3 cross-attention using the public
extension mechanism — `ModelPatcher.add_object_patch()` on each block's cross-attn `forward`, plus
`transformer_options["optimized_attention_override"]` to route mask-bearing calls to
`comfy.ldm.modules.attention.attention_pytorch`. This works on **ComfyUI v0.22.0** (~500s) but **hangs indefinitely on v0.23.0**:
sampling sits at `0/N` with the GPU pinned at 100% and never advances. Reverting ComfyUI to v0.22.0 fixes it completely.

Filed downstream at kijai/ComfyUI-PromptRelay#18; opening here in case it's a core-side regression in the object_patch /
attention-override / masked-attention path.

### Actual Behavior

## Environment
- ComfyUI: **0.23.0** (deps: comfy-aimdo 0.4.7, comfy-kitchen 0.2.10, frontend 1.44.19)
- OS: Windows 11
- GPU: RTX 5090 32GB (Blackwell), driver 595.79
- Python 3.11.8, PyTorch 2.10.0+cu128, CUDA 12.8
- Attention: SageAttention 2.2.0.post4, launched with `--use-sage-attention --disable-dynamic-vram`
- Model: LTX-2.3 (LTXAV) I2V, 1920x1088, 511 pixel frames (64 latent frames), 8 steps

### Steps to Reproduce

## Symptom (log)
Models load fine, then:
[PromptRelay] Built penalty matrix (scaled): Lq=34170, Lk=1024, nonzero=13041243/34990080
[PromptRelay] Built penalty matrix (scaled): Lq=426, Lk=1024, nonzero=162567/436224
0%| | 0/8 [00:00