Fix HuBERT positional batch norm padding handling

[kunwar-vikrant]

What does this PR do?

When conv_pos_batch_norm=True, HuBERT includes padded time steps in BatchNorm statistics. BatchNorm may also turn zero-padded positions into non-zero values, which then affect valid boundary tokens through the positional convolution.

• Passes the original 2D attention mask to HubertPositionalConvEmbedding.
• Excludes padded positions from BatchNorm statistics during training.
• Zeros padded positions before the positional convolution.
• Supports both standard and stable-layer-norm HuBERT encoders.
• Makes the changes in modular_hubert.py and regenerates modeling_hubert.py.
• Adds regression coverage for both training and evaluation modes.

Tests

• HuBERT test suite: 188 passed, 195 skipped.
• Modular conversion consistency check passed.
• Code style checks passed.

Fixes #47739

[github-actions]

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

run-slow: hubert

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 30899056454
Result: success | Grafana metrics are not available yet.

[Abdulrhman-shaheen]

please, read my comment on #47763

[kunwar-vikrant]

@Abdulrhman-shaheen here modeling_hubert.py is generated using modular_hubert.py which is the source of truth.