Low perf for backward pass of dropout fused with normalization

Nobody has claimed this yet.

The fusion definition generated by Thunder for the backward pass of dropout fused with normalization recomputes the dropout mask, resulting in a 1.4× performance degradation compared to a [previous version](https://github.com/NVIDIA/Fuser/blob/df4e918730d77a7f88c1bb3b24625cb2445f58a0/benchmarks/python/test_dropout_rmsnorm_bwd.py#L20) that reused saved masks

To reproduce, use nvFuser main branch run test `test_dropout_rmsnorm_bwd.py` with:
```
pytest /opt/pytorch/nvfuser/benchmarks/python/test_dropout_rmsnorm_bwd.py -vvvs -m 'not skip' -k "float16 and 16384_4096" --benchmark-thunder
```

On B200, output should be:
```
-------------------------------------------------------------------------------------------------------------------------------- benchmark: 2 tests -------------------------------------------------------------------------------------------------------------------------------
Name (time in us) Min Max Mean StdDev Median IQR Outliers OPS (Kops/s) Rounds Iterations
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_dropout_rmsnorm_bwd_nvf_benchmark[dtype=torch.bfloat16-size=[16384_4096]] 151.7090 (1.0) 155.0710 (1.0) 152.6755 (1.0) 1.0656 (6.80) 152.1580 (1.0) 1.2790 (5.63) 2;0 6.5498 (1.0) 10 1
test_dropout_rmsnorm_bwd_baseline_benchmark[dtype=torch.bfloat16-size=[16384_4096]-executor='thunder'] 209.3080 (1.38) 209.7900 (1.35) 209.5808 (1.37) 0.1567 (1.0) 209.6120 (1.38) 0.2270 (1.0) 3;0 4.7714 (0.73) 10 1
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
```

`test_dropout_rmsnorm_bwd_nvf_benchmark` uses saved mask instead of computing dropout achieving about 63% SOL
`test_dropout_rmsnorm_bwd_baseline_benchmark` still computes dropout achieving about 46% SOL

cc @tfogal @riccardofelluga

Start with benchmarks/python/test_dropout_rmsnorm_bwd.py and run the reported pytest command for the float16, 16384_4096 case. Compare test_dropout_rmsnorm_bwd_nvf_benchmark with test_dropout_rmsnorm_bwd_baseline_benchmark and inspect the Thunder-generated backward fusion definition. Done means the backward path reuses the saved mask and approaches the faster benchmark result rather than recomputing dropout.

Written by the indexing model from the issue text.