Add get_sequence_transform to Accuracy for seq2seq models

[joemunene-by]

Summary

Adds Accuracy.get_sequence_transform(pad_index, output_transform) — a static method that returns a pre-built output transform for sequence-to-sequence models, addressing the common NLP/Transformer boilerplate discussed in #3593.

Before (manual boilerplate every project):

def accuracy_transform(output):
    preds, y = output
    preds_flat = preds.reshape(-1, preds.size(-1))
    y_flat = y.reshape(-1)
    mask = y_flat != pad_idx
    return preds_flat[mask], y_flat[mask]

acc = Accuracy(output_transform=accuracy_transform)

After (one-liner):

acc = Accuracy(output_transform=Accuracy.get_sequence_transform(pad_index=0))

What it does

• Flattens (batch, seq_len, num_classes) predictions → (batch * seq_len, num_classes)
• Flattens (batch, seq_len) targets → (batch * seq_len,)
• Masks out positions where target == pad_index
• Supports optional output_transform composition (e.g., extracting from dicts)

Design

Follows @vfdev-5's suggestion in #3593 to use a static method returning a transform rather than creating a new metric class. This keeps the API clean and avoids duplication.

Tests

6 new tests covering:

• Basic padding mask behavior
• No padding (all tokens unmasked)
• End-to-end with Accuracy metric (correct predictions)
• Partial accuracy with wrong predictions
• Custom pad index
• Composition with custom output_transform

All existing tests pass (140 passed, 292 skipped for CUDA/MPS).

Closes #3593

[aaishwarymishra]

Thank you for pr, sorry it took some time reaching out to you.
You see there is already a pr open for this issue #3637 , you can contact the author and ask if they are still working on it or not.

[joemunene-by]

Thanks for the heads up @aaishwarymishra! I wasn't aware of #3637 when I opened this. I can see it's still active with recent commits and review feedback, so happy to close this PR if @Manimaran-tech is continuing with it.

If for any reason they're no longer working on it, I'm glad to pick it up or help finish it. Just let me know!

[aaishwarymishra]

@Manimaran-tech you want to continue on the issue #3593 ?