With multiple scorers, if a scorer fails, GridSearchCV will fail

**What happened**:
I was running GridSearchCV with multiple scoring metrics. One of them ("neg_mean_poisson_deviance") was undefined for some folds b/c it is undefined when y_hat is 0. This was handled during scoring but when `create_cv_results` was called, this raised a `TypeError: 'float' object is not subscriptable`. This is b/c `score` would normally return a dictionary when mutliple scorers are requested but in this case it returned the value I had passed as `error_score` to `GridSearchCV`, which in this case was `np.nan`. The issue is between L274 and L297 in [methods.py](https://github.com/dask/dask-ml/blob/0ea276da1d78db582f40e1c256dfca4f70e6cbc6/dask_ml/model_selection/methods.py#L274).

**What you expected to happen**:
I expected that score to be `np.nan` for the folds in which the scorer failed, but not to raise an error

**Minimal Complete Verifiable Example**:
```python
from sklearn.linear_model import LinearRegression
from dask_ml.model_selection import GridSearchCV
from sklearn.model_selection import LeaveOneOut
import numpy as np

X = np.array([[1, 2],
[2, 1],
[0, 0]])

y = 3 * X[:, 0] + 4 * X[:, 1]
cv = LeaveOneOut()

ols = LinearRegression(fit_intercept=False)
regr = GridSearchCV(
ols,
{"normalize": [False, True]},
scoring=["neg_mean_squared_error", "neg_mean_poisson_deviance"],
refit=False,
cv=cv,
error_score=np.nan,
n_jobs=1
)
regr.fit(X,y)
```

This gives the TypeError I mentioned

**Anything else we need to know?**:
I think this *should* be a fairly quick fix so I'm going to give it a try