Add LTXAV model detection monkey-patch for checkpoint compatibility

[Marcel (goodguy1963)]

Problem

Loading LTXAV GGUF checkpoints fails with RuntimeError: size mismatch because the ComfyUI core model detection doesn't auto-detect these parameters from the checkpoint:

• cross_attention_adaln — checkpoint uses 9 ada params per block, core defaults to 6
• Audio cross-attention dim & head dim — not detected from audio_attn2 weights
• Embeddings connector dimensions — defaults to 30 heads × 128 dim = 3840, but checkpoints use 32 heads with video=4096 and audio=2048

Fix

model_detection_patch.py — a zero-touch monkey-patch applied at ComfyUI-LTXVideo import time that wraps comfy.model_detection.detect_unet_config. When an LTXAV model is detected, it reads the missing parameters from the state dict keys:

Parameter	Detection method
cross_attention_adaln	prompt_scale_shift_table key existence
audio_cross_attention_dim	audio_attn2.to_k.weight shape[1]
audio_attention_head_dim	audio_attn2.to_k.weight shape[0] ÷ 32
Connector dims	learnable_registers shape[1] ÷ 32

No core ComfyUI files are modified — the patch lives entirely within the custom node and activates via a one-line import in __init__.py.

Testing

Verified model creation produces exact checkpoint-matching shapes:

Block 0: scale_shift=[9, 4096] ✓, audio_scale_shift=[9, 2048] ✓
Video connector register: [128, 4096] ✓
Audio connector register: [128, 2048] ✓