[bug] Performance different between sklearn and dask-ml LogisticRegression for multiple classification problem

I tried to load the trained model from dask-ml into sklearn. It gave different results.
And I found out that, even I use fresh trained model, dask-ml estimator's result is still different with sklearn esitmator.

Here is the code I used:

```
from dask_ml.linear_model import LogisticRegression as DaskLogisticRegression
from sklearn.linear_model import LogisticRegression
dask_m = DaskLogisticRegression(solver='lbfgs')
dask_m.fit(X_train.values, y_train.values)
predict = dask_m.predict(X_test.values)
test_acc = metrics.accuracy_score(y_test, predict)
print('test acc', test_acc)

m = LogisticRegression(multi_class='ovr', n_jobs=1)
m.fit(X_train.values, y_train.values)
predict = m.predict(X_test.values)
test_acc = metrics.accuracy_score(y_test, predict)
print('test acc', test_acc)

```
output:
```
test acc 0.1609371578716882
test acc 0.4484344208451938
```

I tried to make two models have the same parameters. However, the performance results are different.
Even if I use the empty/default parameters setting, the performance results are still different.
Any idea?

The data I used has >2 classes.
When I tried it with a binary classification problem, the result is very close. Therefore, the problem happens for multiple classes.