remove index_fill side-effect for scalar tensors (#52209)
Summary:
`index_fill` silently promotes zero dim Tensors to 1-dim Tensors. This PR fixes that.
Was:
```
In [1]: import torch
In [2]: x = torch.tensor(1)
In [3]: idx = torch.tensor(0).long()
In [4]: x.dim()
Out[4]: 0
In [5]: x.index_fill(0, idx, -1).dim()
Out[5]: 1
```
Now:
```
In [6]: x.index_fill(0, idx, -1).dim()
Out[6]: 0
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52209
Reviewed By: ejguan
Differential Revision: D26446470
Pulled By: ngimel
fbshipit-source-id: 4737e6941a7216b57f3416b59362817834df3a3a
