Avoid compute in datasets

https://github.com/dask/dask-ml/blob/d5801584d092d8f13f1b38aaf4da5dc3caa6a213/dask_ml/datasets.py#L332 isn't great, especially in settings like Hyperband #221, that are using the distributed scheduler.

We could probably replace

```python
rng = dask_ml.utils.check_random_state(random_state)

```

with

```python
rng = sklearn.utils.check_random_state(random_state)
```

and draw

1. `informative_idx`
2. random data to seed the `dask.array.RandomState` that is eventually used to generate the large random data.