Remove the hash-merge utility and switch it to use mainline dask.merge

[VibhuJawa]

Remove the hash-merge utility and switch it to use mainline dask.merge

We added the hash-merge utility because of the difference of implementation b/w dask_cudf repartition and the shuffle function being used by dask's merge.

The earlier dask.dataframe shuffle implementation used a different code than the dask_cudf and was more memory-hungry.

We have upstreamed our repartition function to dask-mainline since then, and it should now have the same performance characteristics as this merge, so we should use that instead after we scale test to verify similar results.