add torchvision: Qwen3-VL processor imports it at load time

requirements.txt

@@ -1,6 +1,10 @@
 # Pinned against the ZeroGPU-supported matrix and the Qwen3-VL model card.
 # ZeroGPU supports torch 2.8.0+; the model card specifies torch==2.8.0.
 torch==2.8.0
+# Required, not optional: Qwen3-VL's processor pulls in Qwen3VLVideoProcessor,
+# which imports torchvision at load time. Without it the model fails to load on
+# every backend. 0.23.x is the torchvision paired with torch 2.8.
+torchvision==0.23.0
 transformers>=4.57.0
 sentence-transformers>=5.0.0
 qwen-vl-utils>=0.0.14