datetime operations on NaT produce dangerous results 

```
In [189]: ser = pd.Series([np.datetime64('nat')])

In [190]: ser
Out[190]: 
0   NaT
dtype: datetime64[ns]

In [191]: ser.dt.hour
Out[191]: 
0   -1
dtype: int64
```

While -1 is not a valid hour it is a very valid integer and could (has) caused hard to track bugs. 
