Fix flipped (height,width) in batched inference notebook

[Vision-Kek]

(height,width) is flipped in add_visual_prompt and add_text_prompt in examples/sam3_image_batched_inference.ipynb:

• the size of a SAMImage is h, w, not w, h
• InferenceMetadata also expects (height, width)

Since it's flipped twice, it doesn't affect behavior in the notebook, but w misleadingly stores the height. So either

• change to h, w = datapoint.images[0].size (current proposal),
• or, change to w, h = datapoint.images[0].data.size (which would access the PIL Image which is w,h)

The other changed line 84 regarding torch.inference_mode has been concurrently addressed by #245.