[bug] kNumExpertsPerSM = 4 cause hung on H20

Hello,

Recently, we are trying to run deepep on some H20s and it always hung after some steps with recent commit (no such issue on older version, like before June 2nd). So we revert the commit one by one and reproduce this issue. Finally, we found it's https://github.com/deepseek-ai/DeepEP/commit/d4f3497 which introduced the issue:
`constexpr int kNumThreads = 256, kNumExpertsPerSM = 32, kNumRanksPerSM = 8;` => `constexpr int kNumThreads = 256, kNumExpertsPerSM = 4, kNumRanksPerSM = 8;`

Not sure why this happen. Anyone hit this issue?

Thanks,