fix(engine): use model-native packed attention masks

[wm19999]

Description

Remove the model-family allowlist that injected a hard-coded per-layer attention-mask dictionary into FSDP packed microbatches. Passing None lets each Transformers model build the backend-specific mask it expects from AReaL's reset position IDs, avoiding the Llama/MiniCPM5 dict.ndim crash while keeping Qwen and multimodal preparation model-native.

This also adds CPU regressions for:

• Llama/MiniCPM5-style, Qwen2, Qwen3, Qwen3-MoE, Qwen3.5, and Gemma3 configs
• Qwen2.5-VL and Qwen3-VL mRoPE preparation
• Packed-sequence isolation under eager attention and SDPA

Related Issue

Fixes #1557

Related to #1132, #1153, and #1442.

Type of Change

• 🐛 Bug fix
• ✨ New feature
• 💥 Breaking change
• 📝 Documentation update
• ♻️ Refactoring
• ⚡ Performance improvement
• ✅ Test coverage improvement

Checklist

• I have read the Contributing Guide
• Pre-commit hooks pass (pre-commit run --all-files)
• Relevant tests pass; new tests added for new functionality
• Documentation updated (not applicable; no user-facing API or configuration changes)
• Branch is up to date with main
• Self-reviewed via /review-pr command
• This PR was created by a coding agent via /create-pr
• This PR is a breaking change

Breaking Change Details (if applicable):

None.

Additional Context

Validation:

• PYTHONPATH=. python -m pytest -q tests/test_fsdp_packed_attention_mask.py — 10 passed
• pre-commit run --all-files — passed
• git diff --check — passed

Documentation is not affected because this only changes internal packed-mask plumbing.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[github-actions]

This pull request has been automatically marked as stale because it has not had recent activity within the last 14 days.

Please add a comment or push new commits to keep it active.

Thank you for your contribution!