LinearRegression doesn't return lazy object

LinearRegression requires even longer time than sklearn version and it doesn't return a lazy object:

```
%%time

from sklearn.pipeline import make_pipeline
from dask_ml.linear_model import LinearRegression

size = 1e6

X = dask.array.arange(2*size).reshape(-1,2)
y = dask.array.arange(size).reshape(-1,1)
reg = LinearRegression()
reg.fit(X, y)
```

It looks as a Dask-incompatible function.

Start with the LinearRegression.fit entry point and reproduce the timing example in the issue using the shown Dask arrays. Compare its behavior with the sklearn version and determine whether fit returns a lazy object. Done means the reported incompatibility and excessive runtime are addressed, with regression coverage added where appropriate.

Written by the indexing model from the issue text.