Quick out-of-core downsampling to provide an overview of large data

As suggested in https://github.com/bokeh/datashader/issues/553, we need to improve our raster support so that it can access large rasters chunk by chunk. This will let us work with raster datasets larger than system memory, just as we can currently work with larger-than-memory columnar Dask dataframes, and should make it very practical to work with a small, zoomed-in region of a very large raster dataset.

However, even if full resolution is needed only on deep zooms, very often it is not clear *which* region of the large dataset should be viewed at full resolution, until a zoomed-out overview of the data has been provided to the user. And unfortunately, constructing such an overview using datashader will currently require making a complete pass through the *entire* dataset, reading all of the data into Python in order to aggregate it into a coarse grid.

Even if the Python code were to explicitly downsample the data, accessing only a strided subset of the data from disk, with most file formats and storage technologies the entire dataset will still end up being accessed.

After the above issue is addressed, it would be great if we could work on some specific combinations of aggregations and file formats where constructing such an overview would be relatively cheap. We could also consider making examples of having some pre-processed overview of the data stored with it, and then using HoloViews to switch dynamically between the overview and the full data depending on zoom level.