MiniMax H3 INT8 ConvRot VAE produces black video with NaN/Inf output

## Bug report

Using `minimax_h3_video_vae_int8_convrot.safetensors` with the MiniMax H3 workflow produces an entirely black video. The same workflow works with `minimax_h3_video_vae_fp16.safetensors`.

## Reproduction

- ComfyUI: current local build
- Model: MiniMax H3
- VAE: `minimax_h3_video_vae_int8_convrot.safetensors`
- GPU: RTX 2080 Ti, 22 GB
- Resolution: 864×480
- Length: 362 frames
- Sampling: 4 steps
- VAE path: native H3 spatial and temporal tiling

The output video is completely black. The log also reports:

```text
RuntimeWarning: invalid value encountered in cast
```

This occurs while converting the decoded video tensor to 8-bit output, suggesting that the VAE output contains NaN or Inf values.

Relevant log information:

```text
VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
Requested to load MiniMaxH3VideoVAE
[VAEDecode] shape=(1, 24, 107, 30, 54)
```

The INT8 VAE file was loaded successfully, but the decoded output is invalid.

Could you please check whether this INT8 ConvRot VAE requires a specific ComfyUI, comfy-kitchen, dtype, attention backend, or loading configuration? Is this checkpoint compatible with the current MiniMax H3 VAE loader on RTX 2080 Ti/SM75?

Start by reproducing the MiniMax H3 workflow with minimax_h3_video_vae_int8_convrot.safetensors and compare it with minimax_h3_video_vae_fp16.safetensors. Read the MiniMax H3 VAE loader and VAEDecode path around the reported tensor shape, then check the dtype, attention backend, and loading configuration; done means the INT8 checkpoint decodes finite, non-black video on the tested setup.

Written by the indexing model from the issue text.