Fix EP8 DP/TP RSAG init and empty LM head

[yubofredwang]

Summary

This PR narrows the EP8/DP2/TP4 fix to two concrete runtime issues:

• Precreate the dense TP TritonRSAG state during distributed initialization for attention-DP + dense-supergroup topologies. This avoids divergent first-use ordering of RSAG state creation across ranks before the hidden/logits boundary is reached.
• Treat an empty local hidden batch as valid at the LM-head boundary. When DP attention leaves a rank with zero local tokens, the logits processor now returns an empty logits tensor for local-logits/single-TP paths instead of launching the Kimi fused LM-head kernel.

Follow-up work is intentionally left out of this PR: global-to-local gather_ids layout handling, sampler output merge semantics, output logprobs with local logits, and EAGLE3/NextN hidden-capture merging.

Test Plan

• .venv/bin/python -m pytest test/runtime/distributed/test_comm_ops.py::TestAutoBackendRsagPrecreate test/runtime/test_logits_processor.py -q
• .venv/bin/pre-commit run --all-files

Additional manual validation from the investigation:

• Direct EP8/DP2/TP4 engine smoke passed with the RSAG precreate fix.
• Kimi-K2.5 NVFP4 AIME benchmark passed with score above threshold.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 16db5f6e86

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".