ltx-2 OOMs still with --normalvram --reserve-vram=0 on a dedicated GPU, we expect no OOM

## Bug

LTX-2 19B (fp8) runs out of VRAM with `normalvram=True` on a 24GB RTX 3090. The OOM occurs during text encoding in the `audio_embeddings_connector` forward pass.

This is on master.

## Environment

- GPU: NVIDIA RTX 3090 (24 GB)
- PyTorch: latest
- ComfyUI: 0.10.0

## Reproduction

Run the template ltx2 workflow

## Test Results

| Configuration | Result | Peak VRAM | Duration |
|:---|:---|:---|:---|
| `novram=True` | PASS | 6.74 GB | 211s |
| `normalvram=True`, `reserve-vram=0` | **FAIL (OOM)** | 21.90 GB | 34s |

## Stack Trace

```
File "comfy/text_encoders/lt.py", in encode_token_weights
out_audio = self.audio_embeddings_connector(out)[0]
...
File "comfy/ops.py" in cast_bias_weight
weight = weight.to(dtype=dtype)
torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 226.00 MiB.
GPU 0 has a total capacity of 22.06 GiB of which 112.69 MiB is free.
```

## Workaround

Use `--novram` when running LTX-2.