Experiment with calling `dataset.load()` _after_ joining examples into batch

For the Zarr DataSources, it may be faster to load the data into memory _after_ joining (lazily loaded) examples.

i.e. call `.load()` towards the end of `get_batch()` instead of at the end of `get_example()`.

This should allow `dask` to do a better job of scheduling what needs to be done.  And might result in faster times per batch.
