Allow working with larger-than-memory rasters

[tomalrussell]

E.g. global ~90m map may be on the order of 100GBs if represented uncompressed in memory.

snail.intersection.get_raster_values_for_splits currently takes raster data as a 2D ndarray.

When attributing values from rasters to split and indexed geometries, find an approach which doesn't assume that the full raster array can be loaded in memory.