Fix metric tests

[sananakther6642]

Summary:

• Fix flaky MulticlassAccuracy test when a class has zero samples by explicitly handling empty classes.
• Fix the same zero-sample-class edge case in functional multiclass_accuracy tests.
• Fix BinaryAUROC test failures caused by float64 vs float32 dtype mismatches.
• Skip FrechetAudioDistance VGGish tests when torchaudio prototype features are unavailable.

Test plan:

• pytest -q torcheval/tests/metrics/classification/test_accuracy.py
• pytest -q torcheval/tests/metrics/functional/classification/test_accuracy.py
• pytest -q torcheval/tests/metrics/classification/test_auroc.py
• pytest -q torcheval/tests/metrics/audio/test_fad.py
• pytest -q torcheval/tests/metrics/functional/classification

Fixes: N/A

[sananakther6642]

Hi team! 👋 Just wanted to do a gentle follow-up on this PR — it's been a few months since submission and I wanted to check if there's anything on my end I can help with, such as rebasing, addressing feedback, or making any adjustments. Happy to do whatever is needed to move this forward. Thanks for your time and for maintaining this project! 🙏