Inefficient reading of slices of a Dataset

[jjhelmus]

When reading data from a Dataset, pyfive currently loads all chunks into memory before slicing the requested data. This behavior is inefficient when only a small region of the data is required which could be extracted from a small number or even a single chunk.

The code used for slicing dask arrays may be helpful for determining which chunks need to be read for the given slice.

[mrocklin]

I was looking into pyfive to read cloud-hosted data (we have Python file objects for S3, GCS, Azure) and was sad to learn that slicing doesn't happen cleverly.

[jjhelmus]

I believe the indexing code used in zarr could be adapted for use in pyfive to provide more efficient slicing.