partial_fit should raise an error when `y` in each batch contains labels not present in `classes_`

```
X_first = rng.randn(5, 2)
ovr = OneVsRestClassifier(SGDClassifier())
ovr.partial_fit(X_first, [2, 0, 0, 1, 0], [0, 1, 2])
ovr.partial_fit(X_first, [3, 4, 4, 2, 3])
```

This is the problem with the base estimator as well

```
sgd = SGDClassifier()
sgd.partial_fit(X_first, [2, 0, 0, 1, 0], [0, 1, 2])
sgd.partial_fit(X_first, [3, 4, 4, 2, 3])
```
