[Bug] Fix fp8 deepgemm batch invariant

[yewentao256]

Purpose

Selecting kernels between run_flashinfer_deepgemm_swapAB and run_deepgemm will break batch invaraince, this PR fixes the issue.

Test

VLLM_TEST_MODEL="Qwen/Qwen3-30B-A3B-Thinking-2507-FP8" pytest tests/v1/determinism/test_batch_invariance.py -svx

# now
============== 10 passed, 28 warnings in 458.02s (0:07:38) ===============
# original
======================= short test summary info ========================
FAILED tests/v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[FLASH_ATTN] - Failed: Nondeterministic outputs detected: 1 failed out of 1 trials...
!!!!!!!!!!!!!!!!!!!!!! stopping after 1 failures !!!!!!!!!!!!!!!!!!!!!!!
============== 1 failed, 19 warnings in 79.75s (0:01:19) ===============
sys:1: DeprecationWarning: builtin type swigvarlink has no __module__ attribute

[gemini-code-assist]

Code Review

This pull request addresses a batch invariance issue in the FP8 deepgemm implementation by modifying the kernel selection logic. When batch invariance mode is active, the code now bypasses the dynamic selection between run_flashinfer_deepgemm_swapAB and run_deepgemm, and instead consistently calls run_deepgemm. This change ensures deterministic outputs across different batch sizes. The implementation is straightforward and appears to correctly resolve the described issue.