Fix vision model loading: remove device_map=auto for MiniCPM-V

The MiniCPMV class lacks the all_tied_weights_keys attribute that
device_map=auto requires via infer_auto_device_map. ZeroGPU handles
GPU placement anyway, so device_map is unnecessary.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

models.py

@@ -154,7 +154,6 @@ if not DEMO_MODE:
         MODEL_VISION,
         torch_dtype=torch.bfloat16,
         trust_remote_code=True,
-        device_map="auto",
     )
 
     print("[LifeLog] Loading image generation model…")