torch.zeros/ones RuntimeError when the number of elements approaches upper limit of int32

Reproducer:
```
import torch
x = torch.ones(2**31 - 1, device=torch.device("xpu:0"), dtype=torch.int8)
// RuntimeError: Non-uniform work-groups are not supported by the target device
```
Overflow occurred in `aten::ceil_div<int>` for calculating the number of work groups.