[Sm90] mxfp8

[KareemMusleh]

Right only the forward is implemented. Depends on this quack PR

Will write a summary tomorrow

[KareemMusleh]

The difference between this and bf16 sonicmoe is that we quant the activations before the up_proj/down_proj. And that mxfp8 scales of A need to be dQaccum padded so that we can load them effectively with TMA. In the forward this happens in two places:

1. Before the up_proj. Group by expert and pad the scales. This is done as a separate kernel gather_padded_sfa.
2. Before the down_proj. We just pad the scales because they are already grouped by expert. This is fused into the quant kernel. Now it's its own kernel.

We pre-compute the metadata needed to do this effectively (padded_gather_idx and padded_grouped_idx).

I'd be glad to listen to your feedback.