Use tree reduction in `da.bincount`

Currently `da.bincount` does a [reduction over `np.bincount`s from each chunk]( https://github.com/dask/dask/blob/2f308e72a46de818055819108429e8f7ce7ba03b/dask/array/routines.py#L549 ). It would be nice to change this to use [a tree reduction]( https://github.com/dask/dask/blob/master/dask/array/reductions.py ) instead.