Allowing HyperbandSearchCV to also work with non dask arrays

I have a custom packages which uses a custom data-structure similar to `pandas DataFrame` but that also carry additional metadata. Let's call it `CustomFrame`.

I wrapped all estimator which share the `sklearn` API such as `StandardScaler`, `Lasso`, `XGBRegressor` etc.. to make them able to handle this `CustomFrame`.

This works well with `dask` implementation of `GridSearchCV` or `RandomizedSearchCV` as they just pass the `X` and `y` to the underlying estimators that, as I said, can correctly handle my `CustomFrame`.

I wanted to do the same with HyperbandSearchCV but I found this much more difficult, if not impossible without rewriting most of it, because the implementation is very different as it relies on Future objects and async await.

To be more precise at the moment it doesn't even work with `pandas DataFrames`. If you don't specify the `test_size`, the earliest time it fails is in `dask_ml.model_selection._incremental` at line 569 `test_size = min(0.2, 1 / X.npartitions)` because `DataFrame` does not have the attribute `npartitions`.

I would like to ask if you could consider adding the possibility to also handle also other `ArrayLike` objects the same way is already possible using `GridSearchCV` or `RandomizedSearchCV`

Many thanks
Gio