fix(attention): unwrap hybrid backend for FP8 DSA MHA

[ActiveSky]

Motivation

Fixes #38477. FP8 DSA MHA reads forward_metadata from the outer
HybridLinearAttnBackend during cached-prefix prefill, although the metadata
is owned by its full-attention backend.

Modifications

• Reuse the existing attention-backend resolver when reading FP8 DSA metadata.
• Add a CPU-only regression covering TBO-wrapped hybrid backends.

Validation

• pre-commit run --files python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py test/registered/unit/models/test_deepseek_fp8_dsa_backend.py passed.
• python -m py_compile passed for the changed Python files.

Checklist

• Format code with the repository pre-commit configuration.
• Add focused regression tests.
• No documentation update is required.
• No speed benchmark is required.

---

CI States

Latest PR Test (Base): ❌ Run #34226402992
Latest PR Test (Extra): ❌ Run #34226402882
Latest PR Test (AMD ROCm 7.2): ❌ Run #34226402994

[ActiveSky]

Hi maintainers, thank you for taking a look at this PR.

The implementation and focused checks for #38477 are complete. The current CI gate stops before scheduling the test jobs because this PR does not yet have the required run-ci label. If appropriate, could someone with the necessary permission please add run-ci and rerun the CI?

Thank you for your guidance.

[ActiveSky]

Hi @fzyzcjy @ispobock @ch-wan @Fridge003,

There has been no formal review activity on this PR for some time, so we wanted to check whether this change is still relevant and whether we should continue working on it. We understand priorities can change; if the issue is already covered elsewhere or the scope should be adjusted, please let us know and we will follow your guidance.

If the PR is still useful, could you please take a look when you have time? Feedback on the implementation, tests, or whether the branch should be updated would be very helpful. We hope to contribute in a way that helps move the SGLang project forward.

Thank you very much for your time.