Fix STG crash/index miscount with guide-mask self-attention

[Jean J. de Jong (jjdejong)]

Summary

Fixes a crash (and a related attention-index miscount) in the STG / multimodal
guider whenever cond-image / keyframe guides with strength != 1.0 are combined
with STG perturbation.

Symptom

During the STG-perturbed denoise step:

RuntimeError: The expanded size of the tensor (N) must match the existing
size (M) at non-singleton dimension 1 ... in _attention_with_guide_mask

Root cause

ComfyUI core (CORE-166, "Reduce LTX2.3 peak VRAM when guide_mask is in use")
now splits one video self-attention into up to three optimized_attention
calls over sliced queries against the full key/value. The plugin's STG
PatchAttention skipped a layer by return v (the full-length value) and
counted each sub-call as a separate attention index. So:

• return v is the wrong length for the sliced-query output slot → crash; and
• the extra sub-calls shift audio_attn_idx in calc_stg_indexes, so audio STG
  would skip the wrong attention even when it doesn't crash.

Fix

PatchAttention now recognises a guide-split sub-call via the
low_precision_attention=False kwarg (the only signal core's guide path passes —
this avoids false-positiving the v2a cross-attention, which also has
q_len < v_len), collapses the split into a single logical STG index, and
returns the matching v[:, off:off+q_len] slice when skipping. No core changes.

Scope

Not AV-specific: any video-only workflow combining optional_cond_images
(strength != 1.0) with the STG / multimodal guider triggers this.

Repro

A video-only LTX-2 workflow with optional_cond_images at strength != 1.0 and
the STG / multimodal guider with perturbation enabled.

Caveat

The detector keys on low_precision_attention=False as the guide-split marker,
which is a core implementation detail. If core changes that, the detector would
silently stop collapsing the split — maintainers may prefer a more explicit
contract or a core-side fix.