Perf regression since sha f85ec6

Hi team, I am investigating a regression in performance for `m_grouped_gemm_fp8_fp8_bf16_nt_masked`. Benchmarking script to repro: https://gist.github.com/hj-mistral/d38801ce8e35860a7faba1e1688546cc.

## Env
GPU: H200
CUDA: 12.9

## Script output

### On sha 79f48ee15a82dd5fad5cd9beaa393c1f755e6b55 (current head)
```
Average time per iteration: 26.55 us
Bandwidth: 1030.01 GB/s
```

### On sha ea9c5d9270226c5dd7a577c212e9ea385f6ef048
```
Average time per iteration: 26.70 us
Bandwidth: 1024.26 GB/s
```

### On sha 3254b758e27a5b2f2ae68279314acb1adcb6c1bc
```
Average time per iteration: 20.16 us
Bandwidth: 1356.20 GB/s
```

Can you confirm?

Start with the linked benchmarking script and compare m_grouped_gemm_fp8_fp8_bf16_nt_masked at commits 79f48ee, ea9c5d9, and 3254b75 on the stated H200/CUDA 12.9 environment. Confirm whether the reported timing difference is reproducible and identify the change associated with the regression; done means the regression is explained or confirmed with benchmark results.

Written by the indexing model from the issue text.