Errant pre-compute using `xgboost.dask.DaskXGBClassifier` and `RandomizedSearchCV`

With xgboost 1.2 and current master on dask-ml, passing in `DaskXGBClassifier()` to `RandomizedSearchCV` fails due to `fit_and_score` receiving numpy arrays instead of dask arrays.

I'm still trying to track down where the errant compute is happening (very possible it's happening on the `xgboost` side, just wanted to raise this here for awareness)

before `fit_and_score` is called, the graph is updated with the keys for `X` and `y`, but when the graph is executed numpy arrays are retrieved instead. (Could be an errant `client.get` on the xgboost side)

**What happened**:
In the notebook, a bunch of "key has failed" messages -- in the terminal this error repeatedly:

```python
distributed.worker - WARNING - Compute Failed
Function: fit_and_score
args: (DaskXGBClassifier(), , array([[-2.27450463, 2.11465447, 1.22691127, ..., -0.88190765,
0.09848558, 0.41462044],
[ 0.43908491, 0.2100192 , -0.18272626, ..., -1.49498831,
1.07761167, 0.20180108],
[ 0.83336784, 2.68069819, 0.8442648 , ..., -0.77917656,
-0.1129625 , -1.29956418],
...,
[-0.08958046, 1.04839125, 0.62148611, ..., -1.2469329 ,
0.94976797, 0.56911496],
[ 1.08962267, 0.78895839, 1.16558679, ..., 0.71315066,
1.04262422, -0.97562332],
[-0.29380091, -1.05777127, 0.6632089 , ..., -1.2499148 ,
0.1178788 , 0.22684799]]), array([0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 0, 0, 0, 1, 0, 1,
0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 0, 0, 1, 0, 0, 1, 1, 0, 1, 0, 1, 0,
0, 0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1,
1, 0, 0, 0, 0, 1, 0, 1, 1, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 1,

kwargs: {}
Exception: TypeError("Expecting or . Got ")
```

**What you expected to happen**:
I expect to get back a trained booster.

**Minimal Complete Verifiable Example**:

```python
from distributed.client import Client
from dask_ml.model_selection import RandomizedSearchCV
from dask_ml.datasets import make_classification
from xgboost.dask import DaskXGBClassifier
import xgboost
import dask_ml
​
print(f"{xgboost.__version__=}")
print(f"{dask_ml.__version__=}")
​
c = Client()
​
param_distibutions = {
"max_depth": [5],
"min_child_weight": [10],
"learning_rate": [0.05],
}
​
X, y = make_classification(n_samples=1000, n_features=20, chunks=(100, 20))
​
estimator = DaskXGBClassifier()
​
estimator
​
clf = RandomizedSearchCV(estimator, param_distibutions)
​
clf.fit(X, y)

xgboost.__version__='1.2.0'
dask_ml.__version__='1.7.1.dev3+gc55c1898.d20201119'
/Users/vjs275/miniforge3/envs/msrm2/lib/python3.8/site-packages/sklearn/model_selection/_search.py:278: UserWarning: The total space of parameters 1 is smaller than n_iter=10. Running 1 iterations. For exhaustive searches, use GridSearchCV.
warnings.warn(
('daskxgbclassifier-fit-score-61fb6731d022ad077e150ddca561ff0e', 0, 0) has failed... retrying
('daskxgbclassifier-fit-score-61fb6731d022ad077e150ddca561ff0e', 0, 0) has failed... retrying
('daskxgbclassifier-fit-score-61fb6731d022ad077e150ddca561ff0e', 0, 0) has failed... retrying
('daskxgbclassifier-fit-score-61fb6731d022ad077e150ddca561ff0e', 0, 0) has failed... retrying
('daskxgbclassifier-fit-score-61fb6731d022ad077e150ddca561ff0e', 0, 2) has failed... retrying
('daskxgbclassifier-fit-score-61fb6731d022ad077e150ddca561ff0e', 0, 0) has failed... retrying
```

**Anything else we need to know?**:
I ran into this with xgboost 1.3 snapshots, but ran this MCVE with xgboost 1.2 to confirm I didn't do anything weird with the snapshot install.

**Environment**:

- Dask version: 2.30.0
- Python version: 3.8.5
- Operating System: OSX
- Install method (conda, pip, source): `conda`

Start by running the minimal complete verifiable example with RandomizedSearchCV and DaskXGBClassifier. Trace the inputs passed to fit_and_score after the task is submitted, checking where X and y become NumPy arrays instead of Dask arrays. Done means clf.fit(X, y) completes and returns a trained booster without the worker TypeError.

Written by the indexing model from the issue text.