Should Incremental.partial_fit be concrete when handed numpy arrays?

What should happen here?

```python
sgd = skleran.SGDClassifier()
inc = Incremental(sgd)

X = np.array(...)
y = np.array(...)
inc.partial_fit(X, y)
```

Everything is concrete. Should we build things up lazily or should we execute directly using `inc.estimator.partial_fit(X, y)`?

Start with the Incremental.partial_fit entry point and compare it with inc.estimator.partial_fit(X, y) for the concrete NumPy-array example in the issue. Resolve whether this path should remain lazy or execute directly, then define and test the agreed behavior for concrete inputs.

Written by the indexing model from the issue text.