Add flash-attn support for H100 acceleration

- Add flash-attn>=2.5.0 for faster attention computation
- Pre-built wheels available for CUDA 12.1 + PyTorch 2.5
- Optimized for H100 Hopper architecture on HF Spaces
- Will significantly speed up Optimus (ViT) inference

requirements.txt

@@ -1,6 +1,7 @@
 --extra-index-url https://download.pytorch.org/whl/cu121
 torch>=2.0.0,<2.6
 torchvision>=0.15.0
+flash-attn>=2.5.0
 open-clip-torch
 gradio>=5.49.0
 loguru>=0.7.3