nvfuser opinfo patch test_nvfuser_extremal_values_native_layer_norm_cuda (#80440)
1. patching override tolerance on `test_nvfuser_correctness_nn_functional_gaussian_nll_loss_cuda_bfloat16` for bfloat16 on A100.
2. `native_layer_norm` returns `nan` mean stats on inputs with extremal `inf`, which differs from codegen outputs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/80440
Approved by: https://github.com/davidberard98
