Hyperband: training for a longer time

Hyperband's rule of thumb works great when I want to sample more parameters than training epochs:

``` python
>>> X, y = make_classification(n_samples=1000)
>>> n_params = 200
>>> n_examples = len(X) * 200
>>> max_iter = n_params
>>> chunksize = n_examples // n_params
>>> chunksize
1000
```

This indicates the model should see the entire dataset every `partial_fit` call:

``` python
>>> from dask_ml.model_selection import HyperbandSearchCV
>>> search = HyperbandSearchCV(model, params, max_iter=200)
>>> search.metadata["n_models"]
143
```

How do I train for longer while still sampling (about) 200 parameters?

---

The rule of thumb won't work because it specifies a chunk size larger than the array. Right now, the only way is to sample more parameters:

``` python
>>> from dask_ml.model_selection import HyperbandSearchCV
>>> search = HyperbandSearchCV(model, params, max_iter=400) # to train for no more than 400 epochs
>>> search.metadata["n_models"]
415
```

That's more hyperparameters than I need to sample.

Start with the HyperbandSearchCV entry point and reproduce the issue's max_iter=200 and max_iter=400 examples. Trace how max_iter affects training duration and metadata["n_models"]; done means allowing longer training while sampling about 200 parameters, with coverage for the requested behavior.

Written by the indexing model from the issue text.