ENH: `df.astype` could accept a dict of {col: type}

This would be consistent with other pandas methods.

The reason I'm running into it is having some NaNs scattered across some `int` and `bool` columns, which converts to float / objects. If I discard those NaNs, it would be nice to do

`df = df.astype({'my_bool', 'bool', 'my_int': 'int'})`

instead of

``` python
df['my_bool'] = df.my_bool.astype('bool')
df['my_int'] = df.my_int.astype('int')
```
