torch.full is inconsistent with np.full

## 🐛 Bug
https://docs.scipy.org/doc/numpy/reference/generated/numpy.full.html

Numpy does type inference on `full` but pytorch does not.

## To reproduce

```
import torch
import numpy as np
print(torch.full((10, 1), 5).dtype)
print(np.full((10, 1), 5).dtype)
```
Gives:
```
torch.float32
int64
```
cc @ezyang @SsnL @gchanan 