[CI] Use a torch-nightly-compatible version of torchaudio

[justinxzhao]

Notes:

• Torch nightly is now running 2.2+.
• We unpin nightly torchaudio==2.1, which is incompatible with torch nightly, v2.2.
• Based on torchaudio's dispatcher migration, audio IO should move to using soundfile instead of torchaudio.save/read when dealing with audio files.
• We skip test_audio_feature.py::test_add_feature_data, test_training_determinism_local_backend and test_build_synthetic_dataset for torch nightly. The first item requires a deeper change to our custom implementation of stft (maybe this can be replaced this with the torch-native implementation).
• Filed Unskip torchaudio-related tests that are incompatible with torch 2.2.0 #3645.

[github-actions]

Unit Test Results

  6 files  +  2    6 suites  +2   40m 31s ⏱️ + 6m 53s
30 tests  -   1  25 ✔️  -   1    5 💤 ±0  0 ❌ ±0 
79 runs  +17  63 ✔️ +11  16 💤 +6  0 ❌ ±0 

Results for commit 4d565b6. ± Comparison against base commit c24deca.

This pull request removes 1 test.

tests.regression_tests.model.test_old_models ‑ test_predict_deprecated_model[respiratory]

♻️ This comment has been updated with latest results.