[Qwen3VLMoe] Update expected text in test_small_model_integration_test_expand (#49301)
After #45783 ("Encode multimodal data only once"), the vision encoder now
processes the image once (batch=1) and shares embeddings across beams,
instead of processing a duplicated batch (batch=N beams). This changes
bf16 layer-norm statistics and shifts beam ranking at the margin, producing
a slightly different beam[1] output. Both outputs are semantically correct.
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>