Use an already trained Keras model to predict on lots of data

A common approach is to train on a bit of data and then use that trained model to predict on lots of data. We could do this using ParallelPostFit in dask-ml, or we can use `X.map_blocks` or `df.map_partitions`. In either case we might want to be a bit careful about avoiding repeated serializations costs. For example, in the following case I suspect that we include the serialized model in every task

```python
# maybe bad?
model = load_model()
predictions = X.map_blocks(model.predict)
```

It's probably better to encourage the user to keep the model delayed

```python
# maybe bad?
model = dask.delayed(load_model)()
predictions = X.map_blocks(model.predict)
```

We should also ensure that dask-ml does this correctly, and includes the model as a single task in the graph so that it gets sent around appropriately (cc @TomAugspurger )

I'm also generally curious if a Keras model that lives on the GPU will eventually make its way back onto the GPU when deserializing.