Fix audio resampling: decoding in chunks should return identical samples to decoding in one go

[NicolasHug]

This is a fix for the audio decoder when resampling happens: a resampled stream in consecutive chunks must return exactly the same samples as decoding it in one go. See test_resample_chunked_matches_full for the desired behavior and see the two big Notes I added for details on the solution. In the main lines: we must align the input and output 'grid' for the resampler to output the right samples, and we enforce that through preroll + skipping some samples.

This should address most of #1601 and specifically #1601 (comment)

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/meta-pytorch/torchcodec/1614

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 5 New Failures, 16 Pending

As of commit 2c0db37 with merge base 1f36e84 ():

NEW FAILURES - The following jobs have failed:

• Linux CPU x86 / install-and-test (3.10, 4) (gh)
  test/test_decoders.py::TestAudioDecoder::test_resample_chunked_matches_full[16000-0.3333333333333333-asset0]
• Linux CPU x86 / install-and-test (3.10, 5) (gh)
  test/test_decoders.py::TestAudioDecoder::test_resample_chunked_matches_full[16000-0.3333333333333333-asset0]
• Linux CPU x86 / install-and-test (3.10, 6) (gh)
  test/test_decoders.py::TestAudioDecoder::test_resample_chunked_matches_full[16000-0.3333333333333333-asset0]
• Linux CUDA x86 / install-and-test (3.10, 12.6, 4) (gh)
  test/test_decoders.py::TestAudioDecoder::test_resample_chunked_matches_full[16000-0.3333333333333333-asset0]
• Linux CUDA x86 / install-and-test (3.10, 13.0, 4) (gh)
  test/test_decoders.py::TestAudioDecoder::test_resample_chunked_matches_full[16000-0.3333333333333333-asset0]

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[NicolasHug]

There are minor test failures on FFmpeg <= 6, I'm going to ignore them here as they seem to be green on the follow-up PRs. I suspect it's exactly what #1615 fixes.