[Bug] `no kernel image is available for execution on the device` when using sage_hub attention backend on RTX 5090 (Blackwell, sm_120)

### Describe the bug

When enabling the experimental sage_hub attention backend on RTX 5090 (Blackwell architecture, compute capability 12.0) with PyTorch 2.8 + CUDA 12.9, inference fails with CUDA kernel compatibility error:
```bash
Error no kernel image is available for execution on the device at line 73 in file /src/csrc/ops.cu
```

### Reproduction

```python
import torch
from diffusers.pipelines.flux2.pipeline_flux2 import Flux2Pipeline
import numpy as np
pipe = Flux2Pipeline.from_pretrained("models/FLUX.2-dev-bnb-4bit", text_encoder=None, torch_dtype=torch.bfloat16).to("cuda:0")
pipe.transformer.set_attention_backend("sage_hub"). # <- Bug here. When I commented out that line, it worked fine.
# pipe.load_lora_weights("models/flux-dev-inpaint/pytorch_lora_weights.safetensors")

def create_random_pil(size=(512, 512)):
arr = np.random.randint(0, 255, (size[1], size[0], 3), dtype=np.uint8)
return Image.fromarray(arr)

coarse_pil = create_random_pil((512, 512))
garment_pil = create_random_pil((512, 512))
prompt_embeds = torch.randn(1, 256, 15360, dtype=torch.bfloat16, device="cuda:0")

images = pipe(
image=[coarse_pil, garment_pil],
prompt_embeds=prompt_embeds,
height=512,
width=512,
guidance_scale=7.5,
num_inference_steps=30,
generator=torch.Generator("cpu").manual_seed(42),
)
```

### Logs

```shell
... ...(Initialize output information)
0%| | 0/30 [00:00
- Using distributed or parallel set-up in script?:
```

### Who can help?

@yiyixuxu @DN6