`cf.aggregate` sometimes needlessly puts in a "transpose" layer into the Dask graph

[davidhassell]

In v3.16.2, cf.aggregate sometimes needlessly puts in a "transpose" layer into the Dask graph that is a null operation, i.e. it doesn't change the axis order. This can have performance implications (it is slower to inspect and optimise a Dask graph with more layers), and prevents the possibility of active storage operations (see #501).

PR to follow