diffusers

[core] use `swap_tensors` in group offloading where possible

#13751

Open