Really great release, especially in terms of usability for other attention types natively!

Is your feature request related to a problem? Please describe.
AFAIU, currently in Attention Dispatcher, we cannot set one attention type to attn1 and another one to attn2 at the transformer level. In its original implementation, SkyReels-V2 has "_native_cudnn" for its self-attentions and "flash_varlen" or "_flash_varlen_3" for its cross-attentions.

Describe the solution you'd like.
Maybe something like:

pipeline.transformer.set_attention_backend({
	#"attn": "flash",
 "self-attentions or attn1": "_native_cudnn",
 "cross-attentions or attn2": "flash_varlen"
})

Describe alternatives you've considered.

for block in pipeline.transformer.blocks:
 block.attn1.set_attention_backend("_native_cudnn")
 block.attn2.set_attention_backend("flash_varlen")