Fix masked causal loss in TransformerBridge

[emerardd]

Description

Propagate the forward-pass attention_mask into TransformerBridge causal loss so padding transitions do not affect return_type="loss", return_type="both", or per-token loss.

The fix also:

• moves the loss mask to the logits device when needed;
• uses NaN-safe masking for ignored token losses;
• zeros fully masked Native attention pattern entries so left padding cannot poison later layers; and
• corrects the per-token loss return annotation to reflect its seq_len - 1 shape.

Fixes #1607

Type of change

• Bug fix (non-breaking change which fixes an issue)

Screenshots

Not applicable.

Validation

• Regression tests: 4 passed
• Affected Bridge unit tests: 102 passed, 1 skipped
• Changed files: pycln, isort, and Black clean
• git diff --check: clean
• uv run mypy .: success on 422 source files

Checklist

• I have commented my code, particularly in hard-to-understand areas
• Documentation is not required for this internal correctness fix
• My changes generate no new warnings
• I have added tests that prove my fix is effective
• The complete unit-test suite passes locally; only the affected unit-test surface was run
• I have not rewritten tests relating to key interfaces which would affect backward compatibility