Add STA to V1

[jzhang38]

• add fa3 backend
• remove all attn dropout args (no one use dropout in 2025, fa3 do not even support dropout)
• add supported_attn_backend attribute for each model(so text encoder will avoid using STA even if we set the attn backend to STA & WAN will not call STA if we did not figure out a good mask strategy)
• add prefix for dit module, so each attn knows which layer is it located at.
• add sta
• make sure changes does not break other functions (such as Wan)
• make sure there is no lining error.

[SolitaryThinker]

pretty sure hf is just down....