Sync EAGLE verify results across TP ranks

[zihaomu]

Summary

• Move EAGLE verify result synchronization out of the sampling-only branch so greedy/HIP verify results are also broadcast across TP ranks.
• Keep predict, accept_index, and num_correct_drafts aligned before later running-batch updates and collectives.
• Add a CPU regression test that exercises the HIP-forced greedy path with a mocked TP group.

Fixes #28815.

Motivation

On ROCm, EAGLE/NEXTN verify enters the greedy path via _is_hip. The sampling path already broadcasts verify outputs across TP ranks, but the greedy/HIP path returned local verify results. If ranks accept different numbers of draft tokens, later shape-dependent collectives can diverge and hang.

The fix treats TP synchronization as a verify-output invariant rather than a sampling-only behavior.

Testing

• python3 -m pytest test/registered/unit/spec/test_eagle_verify_tp_sync.py -q
• pre-commit run --files python/sglang/srt/speculative/eagle_utils.py test/registered/unit/spec/test_eagle_verify_tp_sync.py
• MI300X/gfx942 ROCm container synthetic 2-rank validation: python3 repro_issue_28815_tp_divergence.py --mode repo-fixed --master-port 29622 --join-timeout-seconds 30

---

CI States

Latest PR Test (Base): ❌ Run #28007911393
Latest PR Test (Extra): ❌ Run #28007911148

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[ronhuafeng]

I did an independent CPU-side validation of this PR head (f98b8edc56).

Checks run:

git diff --check 067987d86^..067987d86
python3 -m py_compile python/sglang/srt/speculative/eagle_utils.py test/registered/unit/spec/test_eagle_verify_tp_sync.py
docker run --rm -v $PWD:/workspace -w /workspace lmsysorg/sglang:dev \
  bash -lc 'PYTHONPATH=python python3 -m pytest test/registered/unit/spec/test_eagle_verify_tp_sync.py -q'

Result: the focused test passed (1 passed). The patch moves the TP broadcast to the common verify-output path, so it covers the _is_hip/greedy branch described in #28815 instead of only the sampling branch. I did not run the full MI300X EAGLE/NEXTN workload here, but the added regression test matches the reported missing-broadcast invariant.

Also checked CI status: lint passed; the current red PR-test entries are gate/skip finish jobs rather than a focused lint/test failure.