[Bug]: AutoDeploy: nano3 TP8 accuracy degradation

[lucaslie]

Investigate and fix TP=8 GSM8K accuracy issue for nano3 model, see #8744 and #8744 (comment)

[greg-kwasniewski1]

After testing on both FP32 Mamba cache and on the latest branch, the issue is resolved. See https://nvidia.slack.com/archives/C08T55LHSG4/p1762885374118039?thread_ts=1762292343.810769&cid=C08T55LHSG4