[Determinism] Add deterministic TE fused loss support

[ZhiyuLi-Nvidia]

Is your feature request related to a problem? Please describe.
#6965

No deterministic TE fused loss support.

Revert the change once we have deterministic TE fused loss support

[zhongbozhu]

Adding context for why there is no determinism:

Transformer Engine's fused_moe_aux_loss, specifically the CUDA forward kernel fused_moe_aux_loss_forward_kernel.

It uses floating-point atomicAdd to accumulate partial loss sums. The accumulation order can vary, causing small differences between repeated runs.

The PR's toggle disables this kernel and the fused aux-loss score computation (fused_compute_score_for_moe_aux_loss). It keeps fused_topk_with_score_function enabled.

I made a super early fix to TE, need more time to follow up: https://github.com/zhongbozhu/TransformerEngine/commits/fix/deterministic-moe-aux-loss/

[ZhiyuLi-Nvidia]

Transformer Engine's fused_moe_aux_loss, specifically the CUDA forward kernel fused_moe_aux_loss_forward_kernel.

Thanks a lot for quick support.
Have we checked backward grad as well?