plot_partial_dependence overrides feature names when passing ax

The above snippet will overwrite the feature names to 0 and 1 even though they where passed at the first call. Not a bug strictly speaking but maybe we can make it more intuitive? CC @thomasjpfan 

```py
disp = plot_partial_dependence(
    gbdt, X, features=[0, 1], feature_names=('blah', 'blop'),
    ax=ax)

plot_partial_dependence(
    other_gbdt, X, features=[0, 1],  ax=disp.axes_))
```