Implement grid concatenation and standardize datatype casting #2762
This PR has two main aims:
Before this PR was applied past both casting and concatenation were limited to columnar data formats, which meant that certain operations could not be applied to gridded data, e.g. a HoloMap collapse. Having a dedicated concat implementation for both columnar and gridded data also allows much more efficient concatenation than what is currently in use by methods like
Implementing concatenation along HoloMap dimensions also means that Dataset.groupby operations are now reversible and fixes HoloMap.collapse.
As a followup to this PR we should provide special handling for dask arrays/dataframes during casting. This requires multiple things: