Previously the joining code put the final data into temporary mesh
data, which meant more complexity because it couldn't use the slightly
higher level APIs more commonly used for transferring data.
The process can be simplified by just merging into the active mesh
directly (though some care must be taken when that is used by
multiple selected objects).
This process avoids iterating over attribute data twice, by processing
values at the same time as copying them. Also some of the inner hot
loops are parallelized. There is more opportunity for trivial multi-
threading in the future.
Mismatched attribute types and domains will now choose the higher
complexity type and domain, which should help avoid information loss.
There was a bug with the existing face set processing which would
modify the source meshes. That is now fixed.
113d91aba8