Fix stale expected values in integration tests (cuda sm_86 / Aug04 regressions) (#48171)

* [AyaVision, Doge] Update/add ("cuda", 8) expected values for integration tests (cuda sm_86)

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

* [Phi3, Gemma4, Janus, Youtu, Qwen3VLMoe] Add/update ("cuda", 8) expected values for integration tests (cuda sm_86)

- Phi3: update 3 generation/static-cache tests with new text completions
- Gemma4 test_model_with_image: update ("cuda", 8) to match actual output
- Janus test_model_text_generation_batched: add ("cuda", 8) entry
- Youtu test_dynamic_cache + test_static_cache: add ("cuda", 8) entry
- Qwen3VLMoe: fix test_small_model_integration_test ("small wild cat"),
  test_small_model_integration_test_batch_wo_image ("wild cat species"),
  and test_small_model_integration_test_expand (second beam result)

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

* [Flava] Update expected loss values after torchvision LANCZOS interpolation change

After the switch to native torchvision LANCZOS interpolation (instead of
PIL fallback), image preprocessing produces different pixel values, causing
different model outputs. Update the expected mmm_text, mmm_image, and total
loss values in FlavaForPreTrainingIntegrationTest::test_inference.

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

* [Gemma4, InternVL, Qwen3VLMoe] Fix stale expected values for cuda sm_86

- gemma4 test_model_with_image: update ("cuda", 8) to "ocean and a blue sky in the background" (confirmed by test_model_with_image_batch)
- internvl test_llama_small_model_integration_interleaved_images_videos: update ("cuda", 8) to "After re-examining the images" variant (matches rocm (9,4) output)
- qwen3_vl_moe test_small_model_integration_test_expand: add **manul** bolding in element 1
- qwen3_vl_moe test_small_model_integration_test_batch_different_resolutions: add "small " before "wild cat species" in element 0

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

* [Gemma4, VideoPrism, LFM2Moe, MinimaxM2, DiffusionGemma] Fix stale expected values (cuda sm_86)

- gemma4 test_model_text_only: update ("cuda", (8, 6)) poem output
- videoprism: update ("cuda", 8) to LANCZOS values matching (None, None)
- lfm2_moe test_model_1a8b_logits: update ("cuda", None) MEANS tensor
- minimax_m2 test_small_model_generation: update ("cuda", 8) text
- minimax_m2 test_small_model_logits_batched: update ("cuda", 8) LEFT_UNPADDED
- diffusion_gemma batched long test: update ("cuda", None) token lists

Values captured from CI artifacts (run 32325484824, A10G sm_86).

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

* [Mistral, MusicgenMelody, Voxtral] Fix stale expected values (cuda sm_86)

- mistral test_model_7b_logits: fix ("cuda", 8) slice value 1.7568→1.7559
- musicgen_melody: update all 8 failing tests with actual output values
  (greedy/sampling/text-prompt/text-audio-prompt, mono and stereo)
- voxtral test_mini_single_turn_audio_only: update EXPECTED_OUTPUT text
- voxtral test_mini_multi_turn_text_and_audio: "farewell address" vs "political speech"

Values captured from CI artifacts (run 32325484824, A10G sm_86).

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

* [Mistral, HiggsAudioV2] Fix stale expected values (cuda sm_86)

- mistral test_speculative_generation: update expected text
  ("ketchup/relish" → "mayonnaise/must")
- higgs_audio_v2: add ("cuda", 8) entries for 4 integration tests
  (test_multi_speaker_smart_voice, test_zero_shot_voice_cloning,
   test_multi_speaker_voice_cloning, test_batched_inference)
  Each test now has hardware-specific expected audio tokens for A10G.

Values captured from CI artifacts (run 32325484824, A10G sm_86).

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

* [StableLm] Fix stale MEAN expected values (cuda sm_86)

Update EXPECTED_MEAN values for two logit integration tests:
- test_model_stablelm_3b_4e1t_logits
- test_model_tiny_random_stablelm_2_logits

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

* Revert 6 model test files already failing on Aug 03 baseline

These models were already failing in single-GPU CI on Aug 03 (day before
the regression date), so they cannot be attributed to the torch 2.13 cu130
switch and should not be included in this PR:
- flava
- higgs_audio_v2
- mistral
- musicgen_melody
- stablelm
- voxtral

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

* Revert internvl change — already correctly fixed by PR #48153

PR #48153 (merged Aug 20) already set the correct ("cuda", 8) expected
value for test_llama_small_model_integration_interleaved_images_videos
based on the current torch 2.13 CI environment. Our earlier capture
had an incorrect value.

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

* [lfm2_moe, minimax_m2, qwen3_vl_moe] Fix stale expected values from CI artifacts

- lfm2_moe: update ("cuda", None) logit slice for test_model_1a8b_logits
- minimax_m2: update ("cuda", 8) right_unpadded logits for test_small_model_logits_batched
- qwen3_vl_moe: remove bold markers around "manul" in test_small_model_integration_test_expand

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

* [qwen3_vl_moe] Fix two more stale expected values in integration tests

- test_small_model_integration_test_batch: update both elements from
  "small wild cat native to...montane regions" to actual CI output
  "wild cat species native to...steppes"
- test_small_model_integration_test_batch_different_resolutions: trim
  trailing " regions" from element 0 (max_new_tokens cutoff changed)

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

* [qwen3_vl_moe] Revert test_batch (not an Aug 4 regression); keep diff_resolutions fix

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

---------

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