I noticed that some join methods have things like

        dsk = {
            (name, i): (apply, merge_chunk, [left_key, right_key], kwargs)
            for i, right_key in enumerate(right.__dask_keys__())
        }

where we're generating a low-level graph that could just be done with map_partitions. Using map_partitions in these scenarios would both speed up graph transmission and allow for blockwise fusion across the operations. Refactoring this simple sorts of graphs should be straightforward.

• single_partition_join
• hash_join's merge_chunk
• stack_partitions should use HighLevelGraph.from_collections instead of merging all of the input graphs

cc @rjzamora @ncclementi @jrbourbeau