fix: restore verl 0.9 FSDP configuration and packed outputs

[nashjojo]

Description

With verl 0.9, FSDP/fsdp2 workers fail when their checkpoint config receives mbridge_config, disabled worker profiling fails on tool_config=None, and packed batches without sequence parallelism fail when output preparation reads a missing pad_size.

Emit mbridge options only for Megatron checkpoints across actor, reference and critic configs; keep a usable empty profiler tool mapping; and initialize packed-input padding to zero before the sequence-parallel branch. Existing Megatron options and nonzero sequence-parallel padding are preserved.

Validation

• 12 CPU regression cases against verl 0.9.0 and PyTorch 2.11.0: real checkpoint/profiler dataclass construction for all three worker roles and fsdp/fsdp2/megatron, a packed-input/output round trip at SP=1, and mocked text/VLM sequence-parallel padding. The CPU output test explicitly disables FlashAttention cross entropy.
• Black, isort, flake8, mypy, Python syntax and git diff --check passed for the changed files.
• No GPU training or distributed integration experiment was rerun. The full distributed test suite remains for CI/maintainer validation.

Checklist

• Focused regression tests and static checks passed
• Test docstrings describe the covered contracts
• Code is ready for review
• Full distributed/GPU test suite (not rerun)

[nashjojo]

/unittest-module-trainer

[github-actions]

unittest: Run #1882

Tests 📝	Passed ✅	Failed ❌	Skipped ⏭️	Pending ⏳	Other ❓	Flaky 🍂	Duration ⏱️
29	26	0	3	0	0	0	1h 16m

🎉 All tests passed!

Github Test Reporter by CTRF 💚