FA4 backend is not supported on XPU

### 🐛 Describe the bug

test/nn/attention/test_fa4.py currently only instantiates CUDA device tests. We want to enable XPU test collection for this file so that the coverage matrix reflects XPU. However, FA4 itself is currently CUDA-specific and is not supported by the XPU backend.

The FA4 tests depend on CUDA-only capability checks and the `flash_attn.cute.interface` package:

```
def _fa4_dependencies_available() -> bool:
if not torch.cuda.is_available():
return False
major, _ = torch.cuda.get_device_capability(torch.cuda.current_device())
if major not in (9, 10):
return False

importlib.import_module("flash_attn.cute.interface")

```

### Versions

Pytorch 2.14

Start with test/nn/attention/test_fa4.py, examining how device tests are instantiated and how _fa4_dependencies_available gates collection. Run the file's test collection with an XPU environment; the work is done when XPU appears in the coverage matrix without attempting the CUDA-only FA4 path.

Written by the indexing model from the issue text.