[wav2vec2] Remove tensor.item and dynamic slicing operations in the loop that cause graph break

[chaojun-zhang]

What does this PR do?

Remove tensor.item and dynamic slicing operations in the loop that cause graph break

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[astachowiczhabana]

@libinta this commit is required with next OH release

[imangohari1]

@chaojun-zhang CC: @astachowiczhabana
Can you please run the following tests with and without these changes so we can make sure these are not introducing regressions? Thanks.

GAUDI2_CI=1 RUN_SLOW=true python -m pytest tests/transformers/tests/models/wav2vec2/ -s -v

GAUDI2_CI=1 RUN_SLOW=true python -m  pytest tests/test_examples.py -v -s -k  test_run_audio_classification_wav2vec2-base_multi_card

GAUDI2_CI=1 RUN_SLOW=true python -m pytest tests/test_examples.py -v -s -k test_run_speech_recognition_ctc_wav2vec2-large-lv60_multi_card

[imangohari1]

@chaojun-zhang CC: @astachowiczhabana Can you please run the following tests with and without these changes so we can make sure these are not introducing regressions? Thanks.

GAUDI2_CI=1 RUN_SLOW=true python -m pytest tests/transformers/tests/models/wav2vec2/ -s -v

GAUDI2_CI=1 RUN_SLOW=true python -m  pytest tests/test_examples.py -v -s -k  test_run_audio_classification_wav2vec2-base_multi_card

GAUDI2_CI=1 RUN_SLOW=true python -m pytest tests/test_examples.py -v -s -k test_run_speech_recognition_ctc_wav2vec2-large-lv60_multi_card

@chaojun-zhang
I ran these tests and one of them is failing compared to main.
GAUDI2_CI=1 RUN_SLOW=true python -m pytest tests/test_examples.py -v -s -k test_run_audio_classification_wav2vec2-base_multi_card
with current changes shows

E   eval_accuracy: 0.7394701086956522 vs 0.715572
E   train_runtime: 201.5022 vs 66.578295
E   train_samples_per_second: 703.581 vs 2827.0517999999997
E   eval_samples_per_second: 308.74 vs 3458.01995

and without shows:

E   eval_accuracy: 0.7338654891304348 vs 0.715572
E   train_runtime: 79.2719 vs 66.578295
E   train_samples_per_second: 2731.957 vs 2827.0517999999997
E   eval_samples_per_second: 2812.678 vs 3458.01995

We need to understand what is casuing the perf desegregation on this.

[jiminha]

@astachowiczhabana Could you also check above?