Request support for Sage Attention

[doogyhatts]

First, thanks for making v5.1!

I noticed my generation time on the RTX6000 Ada for 1344x768 resolution, 49 frames with bf16 model precision, to be around 866-876 seconds.
I did notice that teacache is enabled as well.
And system ram peaked at 81gb. (might be because I switched between bf16 and fp8)

Just wondering if there is a possibility of adding Sage Attention to speed up the generation process, other than having teacache?
Thanks!