you could do it today by manually encoding your max parallelism as N solids and then have two solids on the either side for dividing work out to the N solids and one for fanning back in. You could do that in a

CompositeSolidDefinition

so that your top level dag does not look insane.