[deps] feat: bump veomni pin to 58759e7 (FSDP2 pre-backward hook fix)

[Luosuu]

Summary

Bumps the `veomni` pin from `9c041db` to `58759e7`, which is the squash-merge of ByteDance-Seed/VeOmni#731. That PR restores the FSDP2 pre-backward unshard hook on `lm_head` for every MoE / VL architecture that uses VeOmni's chunk_logprobs path.

Why this matters

Without #731, `loss.backward()` on the actor update of any of the affected architectures crashes inside `chunk_logprobs.backward` with:

```
RuntimeError: setStorage: ... storage of size 0
```

Root cause: the patched forward did `output = MoeCausalLMOutputWithPast(...)` followed by `output.log_probs = …; output.entropy = …` after construction. Those mutations were invisible to `ModelOutput`'s pytree flattening, so FSDP2 never registered the pre-backward hook on `lm_head`'s parent unit, and the saved weight reference into the chunk_logprobs custom autograd `Function` stayed sharded by the time backward fired.

The merged squash commit (`58759e7`) contains both the original `qwen3_5_moe` fix (`#731`, from @di.lu) and the cross-arch extension (`ByteDance-Seed/VeOmni#732`) we stacked on top — combined diff covers `qwen3_moe`, `qwen3_5_moe`, `qwen3_vl_moe`, `qwen3_omni_moe`, `qwen2_vl`, `qwen2_5vl`, `qwen3_vl`, and `qwen3_5` (the 8 architectures that shared the post-construction-mutation anti-pattern).

Smoke verification

Both runs on 1×8B200 with the new pin (verl PR #6184 wiring `return_log_probs=True`):

arch	model	step:1 `rollout_probs_diff_max`	step:2 same	`actor/entropy`	actor update
`deepseek_v3`	Moonlight-16B-A3B-Instruct (GRPO + gsm8k)	0.0	0.0	~0.034	✅
`qwen3_moe`	Qwen3-30B-A3B (GRPO + gsm8k, `total_training_steps=2`)	0.0	0.0	~0.124	✅ (8m47s end-to-end, gsm8k acc 0.454)

Same Qwen3-30B-A3B recipe crashed earlier today on `9c041db` with the storage-of-size-0 error in `chunk_logprobs.backward`; runs to completion now.

Why not duplicate

```
gh pr list --repo verl-project/vexact --state open --search "veomni 58759e7 in:title,body"
gh pr list --repo verl-project/vexact --state open --search "fsdp2 in:title,body"
```

Both empty.

Test plan

• `uv lock` resolves the new pin: `Updated veomni (dynamic) -> v0.1.9a4 (58759e78)`.
• Bitwise rollout-actor alignment + actor update verified end-to-end on 1×8B200 for both `deepseek_v3` (Moonlight) and `qwen3_moe` (Qwen3-30B-A3B). Both show `rollout_probs_diff_max == 0.0` across two consecutive training steps.

AI assistance

This PR was authored with AI assistance (Claude Code). The submitting human reviewed every line and ran the smoke tests above before requesting review.

[gemini-code-assist]

Code Review

This pull request updates the veomni dependency to a newer revision in both pyproject.toml and uv.lock. Feedback indicates that the uv.lock file may have been manually edited or only partially updated, as the top-level revision hash was not changed; it is recommended to run uv lock to ensure the lockfile is fully synchronized.

[gemini-code-assist]

The uv.lock file revision (at the top of the file) does not appear to have been updated in this PR, despite the changes to the veomni pin in pyproject.toml. This suggests the lockfile might have been manually edited or only partially staged. It is recommended to run uv lock and commit the entire resulting uv.lock file to ensure that the revision hash and all dependency metadata are correctly synchronized.