fix mm-chunk-embedding test suite

[ZailiWang]

Motivation

Fix CPU ci failure introduced in #31826 .

Modifications

---

CI States

Latest PR Test (Base): 🚫 Run #30894746752
Latest PR Test (Extra): ❌ Run #30894746421

[ZailiWang]

/tag-run-ci-label

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[ZailiWang]

xeon CI passed: https://github.com/sgl-project/sglang/actions/runs/30522562290/job/90806101798?pr=32895

[mingfeima]

only move to base-a won't fix this issue. The test case test_mm_chunked_embedding_unit.py still fails.

=========================== short test summary info ============================
FAILED registered/chunked_prefill/test_mm_chunked_embedding_unit.py::test_by_item_list_matches_tensor - AssertionError: attention tensor model parallel group is not initialized
FAILED registered/chunked_prefill/test_mm_chunked_embedding_unit.py::test_full_list_matches_tensor - AssertionError: attention tensor model parallel group is not initialized
FAILED registered/chunked_prefill/test_mm_chunked_embedding_unit.py::test_full_matches_by_item - AssertionError: attention tensor model parallel group is not initialized
=================== 3 failed, 2 passed, 2 warnings in 0.72s ====================
.
.
End (4/46):
filename='/home/runner/work/sglang/sglang/test/registered/chunked_prefill/test_mm_chunked_embedding_unit.py', elapsed=12, estimated_time=10.0
.
.


✗ FAILED: /home/runner/work/sglang/sglang/test/registered/chunked_prefill/test_mm_chunked_embedding_unit.py returned exit code 1

[mingfeima]

add a runtime fixture in test case to bypass single process CPU for parallel GPU run.