[Bugfix] Release H3 chunked MP4 workers

[0z5a]

Current A100 validation at ae3f5a9b692b (2×A100 40GB; Torch 2.13.0+cu130, vLLM 0.30.0).

Check	Baseline	Current result	Speedup
Native scoped regression suite	N/A	16 passed at 4552e981c383; entire file tree unchanged	N/A
A100 full-model high-concurrency comparison	Pending	Pending	Pending

Full trained MiniMax-H3 FL2VA decode and native MP4 worker completion under high concurrency remain pending.

Current incremental diff against the forked prerequisite (2 paths).

The instance was externally stopped during the remaining test queue. All completed results above were saved off-instance; unfinished measurements remain pending.

Purpose

Depends on unmerged #7519. This Draft carries that prerequisite in the main diff; review the follow-up in commit 8b93e5d, which changes only two files.

Move lishunyang12/vllm-omni#62 to the upstream repository. Related to #7415 (P5 output pipeline).

Extend the MiniMax-H3 chunked MP4 sink's exception handling across setup, decoding and finalization, releasing its worker on failure before handoff. Successful calls still transfer ownership to the caller. Restore the chunk_value_range = (0.0, 1.0) capability declaration.

Test Plan

python -m pytest tests/diffusion/models/minimax_h3/test_chunked_output_lifecycle.py

Parent: 7ba27060000028a1ab5fc86058271e1fb9f9da29. vLLM 0.29.0, PyTorch 2.13.0+cu130, Python 3.12.3.

Test Result

Existing validation of the identical source tree, before this metadata-only republication:

• Lifecycle regression: 16 passed; unchanged parent: 8 failed / 8 passed.
• Lifecycle plus VAE tiling/chunk suites: 49 passed.
• CUDA/pinned D2H/PyAV integration: 14 passed; four 107-frame MP4 variants byte-identical before/after.
• Pretrained VAE: pixels and MP4 bytes match the reference for three shapes up to 107 frames at 256×448 on one RTX PRO 4000 Blackwell; post-decode failure releases the worker. This used a separate validation-only checkpoint callback extension.
• Applicable pre-commit hooks passed, including mypy and test marks.

Full pretrained-model service E2E and multi-GPU failure recovery are not qualified by these results. DLO shutdown changes remain separate in #7581.

Test and benchmark sources are retained locally; the validation results below refer to those local files.

[hsliuustc0106]

This PR touches vllm_omni/diffusion/, tests/diffusion/, docs/user_guide/, recipes/MiniMaxAI/, tools/minimax_h3/ (49 files). Based on CODEOWNERS coverage of the changed files, the most-related reviewers appear to be:

@wtomin @Bounty-hunter @fhfuih

Could one of you take a look when you get a chance? Thanks!