[Gemma3n] Update integration test expected values for A10G + torch 2.13 (#48035)

* [Gemma3n] Update integration test expected values for A10G + torch 2.13

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

* Revert test_model_4b_multiimage cuda expected value — needs investigation

The multiimage test on A10G produces cow-only description instead of
Chinatown, suggesting the model is not attending to the second image
properly. This is a content regression that needs investigation rather
than just an env-driven numeric drift.

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

* [Gemma3n] Also update test_model_4b_multiimage cuda expected value

Sep24 and Aug16 both produce the same cow-description on A10G,
confirming this is an env difference not a code regression.

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

* Revert test_model_4b_multiimage — capture script used wrong inputs

Our capture script sent cow+australia but the actual test only sends
australia.jpg. The Chinatown expected value is correct for the test.
Need to re-run with correct single-image input before updating.

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

* [Gemma3n] Update test_model_4b_multiimage cuda expected value (correct input)

Running australia.jpg alone (the correct single-image input) on A10G
produces slightly different wording: "Here are the key elements" vs
"some of the key elements", ending with "* **A prominent red" vs "* **A".
Content is still Chinatown — same scene, minor env-driven wording diff.

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

* [Gemma3n] Also update test_generation_beyond_sliding_window cuda expected value

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

---------

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>