Fused amax and scale update for TEv2

Nobody has claimed this yet.

## 🚀 Feature

As in TE, there is performance to be gained if we were to use fused amax and scale updates when using the delayed scaling recipe.

This could be achieved two ways, one would be to use TE's function [`tex.fused_amax_and_scale_update_after_reduction`](https://github.com/NVIDIA/TransformerEngine/blob/cae1c436027cc028ee49c83463141ba67f0adca0/transformer_engine/pytorch/fp8.py#L407-L414). The other way to achieve this would be to trace through `_amax_and_scale_update`, therefore also [`_default_get_amax_and_update_history`](https://github.com/NVIDIA/TransformerEngine/blob/cae1c436027cc028ee49c83463141ba67f0adca0/transformer_engine/pytorch/fp8.py#L675-L687) and [`_default_sf_compute`](https://github.com/NVIDIA/TransformerEngine/blob/cae1c436027cc028ee49c83463141ba67f0adca0/transformer_engine/pytorch/fp8.py#L690-L715). The second option might be challenging because of all the `torch.where` used.

Start by reading transformer_engine/pytorch/fp8.py, especially tex.fused_amax_and_scale_update_after_reduction, _amax_and_scale_update, _default_get_amax_and_update_history, and _default_sf_compute. Compare the two proposed implementation paths and determine how delayed scaling could use a fused update. Done means TEv2 uses fused amax and scale updates without changing the intended scaling behavior.

Written by the indexing model from the issue text.