Rely on numel() == 1 to check if distribution parameters are scalar. (#17503)
Summary:
As discussed here #16952, this PR aims at improving the __repr__ for distribution when the provided parameters are torch.Tensor with only one element.
Currently, __repr__() relies on dim() == 0 leading to the following behaviour :
```
>>> torch.distributions.Normal(torch.tensor([1.0]), torch.tensor([0.1]))
Normal(loc: torch.Size([1]), scale: torch.Size([1]))
```
With this PR, the output looks like the following:
```
>>> torch.distributions.Normal(torch.tensor([1.0]), torch.tensor([0.1]))
Normal(loc: 1.0, scale: 0.10000000149011612)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17503
Differential Revision: D14245439
Pulled By: soumith
fbshipit-source-id: a440998905fd60cf2ac9a94f75706021dd9ce5bf
