use xlarge ZeroGPU (full Blackwell) for VAE-decode headroom

Crash is in VAE-decode conv3d only, after the transformer completes, on
every torch version. On the default 'large' size (half a Blackwell, ~48GB)
the large full-res decode allocation fails and PyTorch's OOM path dies in
the forked process's NVML query, surfacing as the nvmlInit assert. Built-in
VAE tiling drops the decoder's timestep conditioning so it's unsafe here;
request the full 96GB Blackwell instead (2x quota cost).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app.py

@@ -140,7 +140,7 @@ def get_duration(prompt, negative_prompt, input_image_filepath, input_video_file
     else:
         return 60
 
-@spaces.GPU(duration=get_duration)
+@spaces.GPU(duration=get_duration, size="xlarge")
 def generate(prompt, negative_prompt, input_image_filepath=None, input_video_filepath=None,
              height_ui=512, width_ui=704, mode="text-to-video",
              duration_ui=2.0,