feat: Patch SageAttention 3 Node

[Haoming02]

• Add a new Node which patches the model to use sageattn3 on non-first/last blocks and non-first/last steps
• Use PromptServer.instance.send_progress_text to notify the user if sageattn3 was not usable
• Models without block_index (e.g. SDXL) will always use optimized_attention

[isaac-mcfadyen]

Just a small comment (I'm not a Comfy contributor) that most users probably won't know what optimized_attention is.

I assume it's the internal name for non-Sage attention i.e. SDPA? If that's the case, wording it more like Apply Sage Attention 3 to the middle blocks and steps, while using normal attention for the first/last blocks and steps might be more understandable.

[Kosinkadink]

The message can say 'default attention' instead of 'optimized_attention', maybe.

[rattus128]

LGTM