BUG: np.fmin behaves differently on object arrays

`np.minimum` is consistent, `np.fmin` is not

```python
>>> x = np.array([1, np.nan])

>>> np.minimum.reduce(x)
nan
>>> np.minimum.reduce(x.astype(object))
nan

>>> np.fmin.reduce(x)
1.0
>>> np.fmin.reduce(x.astype(object))
nan  # oops
```