Inconsistency between torch.abs() and np.absolute() for complex tensors

```
import torch
>>> a = torch.tensor([1+2j])
tensor([(1.0000 + 2.0000j)], dtype=torch.complex128)
>>> a.abs()
tensor([   (2.2361 + 0.0000j)], dtype=torch.complex128)


import numpy as np
>>> a=np.array([1+2j])
>>> a.dtype
dtype('complex128')
>>> np_res = np.absolute(a)
array([2.23606798])
>>> np_res.dtype
dtype('float64')
```

cc @ezyang @anjali411 @dylanbespalko