[CUDA] indexing on large tensor causes invalid configuration argument

### 🐛 Describe the bug

Reproducer:
```
import torch

mask = torch.randint(0, 20, (4, 87, 1056, 736), device="cuda")
to_apply = torch.tensor([True, False, False, True], device="cuda")
mask[to_apply]
```

### Versions

build on main from today

cc @ptrblck @msaroufim @eqy @jerryzh168