Normalize smoothed KL mean by non-padding targets - #3115

What does this PR do?

Fix the mean reduction in kldiv_loss when label smoothing is enabled. It currently runs loss.sum().mean(), so it returns a sum. Duplicating a batch increases the supposedly averaged loss and gradients. The unsmoothed path uses an actual token mean. The small-smoothing limit is checked on unpadded inputs; the existing difference between its length mask and the smoothed path's pad_idx mask is outside this change.

Divide the summed KL divergence by the number of non-pad_idx targets. Keep all-padding input at zero, and document this reduction explicitly. The smoothing distribution, sum, batchmean, batch, and existing length handling are unchanged. Behavior change: smoothed mean values and gradients are now normalized; callers relying on the old sum should request sum explicitly. The bundled LibriSpeech Transformer configuration uses batchmean and retains its behavior.

On two bundled speech samples through the existing MFCC/CRDNN/attention ASR example, 60 valid target tokens produce a loss of 153.60635 before and 2.56011 after, matching an independent torch.distributions KL reference. With the fix, one public Brain.fit_batch SGD step matches independently calculated updates for all 31 trainable parameters on CPU and MPS. This validates loss scaling, not recognition accuracy.

Validation (Python 3.12.14, torch/torchaudio 2.10.0, macOS arm64):

• Original implementation: 7 new regression failures, 3 passing controls. Fixed implementation: 40 passed, 1 skipped across loss unit tests and module doctests; 11 new tests passed on MPS. Coverage includes classification/sequence inputs, repeated batches, padding and its gradients, all-padding input, independent loss/gradient references, and the small-smoothing NLL limit.
• Full unit suite: 767 passed, 6 skipped, 6 failed. All six failures reproduce with the original upstream loss file: four rescorer tests need uncached models, test_fallback_when_all_fail assumes os.sched_getaffinity, and test_run_shell assumes shell echo -n behavior.
• Existing ASR_seq2seq integration test passed. All configured pre-commit hooks passed. Built sdist/wheel with the declared setuptools backend; the extracted wheel passed the same real-audio CPU training-step check.

CUDA, large-scale training, and online model tests were not run. AI assistance: implementation, tests, validation, review, and PR preparation used OpenAI Codex; no independent human review is claimed.

PR review