Return feature attention masks from Wav2Vec2-derived models

[DebadityaHait]

What does this PR do?

Fixes #25307.

Wav2Vec2-derived base models already compute a feature-vector attention mask for padded inputs but did not expose it in their return-dict outputs. This adds an optional attention_mask field to Wav2Vec2BaseModelOutput and returns the exact reduced mask from Wav2Vec2Model, WavLM, Data2Vec Audio, and Wav2Vec2 Conformer.

The mask aligns with last_hidden_state; return_dict=False tuple layouts are unchanged.

The WavLM, Data2Vec Audio, and Wav2Vec2 Conformer return sites are generated from their modular definitions, so they are included to keep repository modular consistency valid.

Code Agent Policy

This was a human-led contribution. AI assistance was used for implementation and testing; the human submitter reviewed every changed line and validated the behavior locally.

• Coordination: Return updated attention mask from Wav2Vec 2.0 #25307, including the maintainer’s July 24, 2026 invitation to revive this work.
• Related work: closed PR Return effective attention mask in Wav2Vec2BaseModelOutput #25471; this retains only the generated Wav2Vec2-derived output paths required by current repository consistency.
• Validation:
  • python -m pytest tests/models/wav2vec2/test_modeling_wav2vec2.py -q — 211 passed, 199 skipped
  • affected WavLM, Data2Vec Audio, and Wav2Vec2 Conformer suites — 251 passed, 403 skipped
  • exact modular conversion check
  • make style

Before submitting

• This was discussed and approved in Return updated attention mask from Wav2Vec 2.0 #25307.
• Added a regression test for the public reduced-mask output.
• Updated the public output docstring.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: data2vec, wav2vec2, wav2vec2_conformer, wavlm

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 30190618585:2
Result: success | Jobs: 16 | Tests: 174,844 | Failures: 0 | Duration: 16h 27m